concept: subjective-evaluation last_updated: '2026-07-27' paper_count: 180 papers: - id: '1609.03499' published_date: "2016-09-12" entry_date: '2026-07-27' year: 2016 venue: arXiv task: - TTS architecture: - autoregressive-LM relevance: medium evidence_role: - historical_context - core_evidence current_role: foundational method_family: - autoregressive_generation claims: - claim_id: direct_generation_of_raw_audio_waveforms_without_intermediate role: supports claim: Direct generation of raw audio waveforms, without intermediate vocoder parameters, produces substantially higher naturalness than parametric or concatenative synthesis pipelines as judged by human listeners. source: §3.2, Table 1 evidence: On the MOS naturalness test using proprietary Google speech databases, WaveNet (conditioned on linguistic features and log-F0) scored 4.21 on North American English and 4.08 on Mandarin Chinese, compared to 3.86 / 3.47 for HMM-driven concatenative and 3.67 / 3.79 for LSTM-RNN parametric baselines (Table 1). confidence: high relevance: high - claim_id: dilated_causal_convolutions_enable_autoregressive_audio_models_to role: supports claim: Dilated causal convolutions enable autoregressive audio models to achieve receptive fields orders of magnitude larger than standard causal convolutions with comparable computational cost. source: §2.1, Figure 3 evidence: 'The core component is a stack of dilated causal convolutions: dilation doubles with each layer (1, 2, 4, ..., 512), and the pattern is repeated across multiple blocks, yielding a receptive field that grows exponentially with depth rather than linearly.' confidence: high relevance: low - claim_id: a_single_autoregressive_model_conditioned_on_speaker_identity role: supports claim: A single autoregressive model conditioned on speaker identity can represent many voices with shared internal structure, and multi-speaker training improves per-speaker quality relative to single-speaker training. source: §3.1 evidence: In the multi-speaker experiment on VCTK (109 speakers, 44 hours), a single global-conditioning model captured all speakers' characteristics, with validation performance improving over single-speaker training, suggesting that speaker-level representations are shared and mutually reinforcing in the model's internal structure. confidence: high relevance: low - claim_id: receptive_field_size_is_a_binding_constraint_for role: supports claim: 'Receptive field size is a binding constraint for prosodic naturalness: when the receptive field is insufficient to cover phrase-level F0 contours, prosody degrades even when segmental quality remains high.' source: §3.2 evidence: Adding the external F0 conditioning resolved this, showing that receptive field size is a binding constraint for prosodic naturalness independent of model capacity. confidence: high relevance: high - claim_id: autoregressive_raw_waveform_generation_achieves_high_naturalness_at role: supports claim: Autoregressive raw-waveform generation achieves high naturalness at the cost of sequential sample-level inference, creating a fundamental speed-quality trade-off that constrains deployment in real-time applications. source: §4, §3.2 evidence: 'All timestep predictions are computed in parallel during training (ground-truth samples available at all positions), but inference is strictly sequential: one sample at a time, each fed back into the network to predict the next.' confidence: high relevance: high limitations: - Inference is strictly sequential at the sample level, requiring approximately one computation step per generated sample. - At the reported generation rates (roughly 1.5× real-time compute), WaveNet is not suitable for real-time TTS deployment without hardware-specific optimisation or a parallel decoding approximation. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '1703.10135' published_date: "2017-03-29" entry_date: '2026-07-27' year: 2017 venue: arXiv task: - TTS architecture: - transformer-enc-dec relevance: medium evidence_role: - historical_context - core_evidence current_role: foundational method_family: - transformer_encoder_decoder claims: - claim_id: end_to_end_tts_models_trained_from_characters role: supports claim: End-to-end TTS models trained from characters with seq2seq attention can match or exceed production statistical parametric systems in subjective naturalness without hand-engineered linguistic features. source: §5.2, Table 2 evidence: The encoder thus produces robust sequence representations without requiring phoneme-level alignment or a hand-crafted linguistic frontend. confidence: high relevance: high - claim_id: training_stability_and_alignment_quality_in_character_level role: supports claim: Training stability and alignment quality in character-level seq2seq TTS improve substantially when the decoder emits multiple output frames per attention step rather than one. source: §3.3 evidence: 'A key design choice is the reduction factor r: the decoder predicts r non-overlapping mel frames per step rather than one, dividing decoding steps by r, accelerating training convergence, and encouraging the attention to advance more readily through the input sequence (r=2 in the main experiments).' confidence: high relevance: low - claim_id: cbhg_style_encoders_combining_multi_scale_convolution_highway role: supports claim: CBHG-style encoders combining multi-scale convolution, highway networks, and bidirectional recurrence yield more robust text representations than standard RNN encoders, reducing mispronunciation rates. source: §3.2, §5.1 evidence: The encoder centres on a novel CBHG module (1-D Convolution Bank, Highway network, and bidirectional GRU). confidence: high relevance: low - claim_id: post_processing_networks_with_access_to_the_full role: supports claim: Post-processing networks with access to the full decoded sequence improve harmonic structure in predicted spectrograms compared to frame-level-only decoding. source: §3.4, §5.1 evidence: Removing the post-processing net eliminates fine harmonic structure in the output. confidence: high relevance: low limitations: - Evaluated on a single internal speaker in a controlled studio environment. - No multi-speaker, out-of-domain, or noisy-data experiments are reported. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '1712.05884' published_date: "2017-12-16" entry_date: '2026-07-27' year: 2017 venue: arXiv task: - TTS architecture: - transformer-enc-dec relevance: medium evidence_role: - historical_context - core_evidence current_role: foundational method_family: - transformer_encoder_decoder claims: - claim_id: tacotron2_mel_intermediate_bridges_e2e role: supports claim: A compact low-level acoustic intermediate representation can bridge text encoding and neural vocoder synthesis without requiring hand-crafted linguistic features, enabling fully end-to-end neural TTS at human-quality levels. source: §2.1, Table 1 evidence: Tacotron 2 reaches MOS 4.526 against a ground-truth MOS of 4.582, closing nearly the entire gap to human recordings on the internal single-speaker test set. confidence: high relevance: high - claim_id: tacotron2_location_sensitive_attention role: supports claim: Location-sensitive attention, which incorporates cumulative attention weights as a conditioning signal, reduces failure modes such as repetition and omission compared to standard additive attention in autoregressive TTS decoders. source: §2.2 evidence: The paper adopts location-sensitive attention specifically to address repetition/omission failure modes observed with plain additive attention in the original Tacotron. confidence: medium relevance: low - claim_id: tacotron2_prenet_bottleneck_essential role: supports claim: A pre-net information bottleneck in the autoregressive decoder is essential for stable attention alignment during training. source: §2.2 evidence: The two-layer 256-unit pre-net bottleneck is described as necessary for stable attention learning in the decoder design. confidence: medium relevance: low - claim_id: tacotron2_vocoder_train_on_predicted role: supports claim: Training the vocoder on predicted rather than ground-truth intermediate features is necessary because predicted features are systematically oversmoothed; vocoders trained on clean features degrade when exposed to predicted inputs. source: §3.3.1, Table 2 evidence: Training WaveNet on ground-truth spectrograms and then synthesising from predicted (oversmoothed) spectrograms is the worst configuration in the paper's ablation (MOS 4.362 vs. 4.526 for the matched train/inference condition). confidence: high relevance: high - claim_id: tacotron2_mel_shrinks_vocoder_receptive_field role: supports claim: Neural vocoders conditioned on compact mel spectrograms can operate with substantially smaller receptive fields than those conditioned on linguistic features, enabling significant architecture simplification without quality loss. source: §3.3.4, Table 4 evidence: A 12-layer WaveNet with a 10.5 ms receptive field matches the quality of the full 30-layer 256 ms model when conditioned on mel spectrograms. confidence: medium relevance: low limitations: - All experiments use a single proprietary speaker and internal, unreleased dataset, leaving generalisation to multiple speakers, languages, or recording conditions untested. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '1904.02882' published_date: "2019-04-05" entry_date: '2026-07-27' year: 2019 venue: arXiv task: - TTS architecture: [] relevance: medium evidence_role: - core_evidence - infrastructure current_role: influential method_family: [] claims: - claim_id: libritts_24khz_higher_mos_than_16khz role: supports claim: Training TTS models at 24 kHz rather than 16 kHz produces measurably higher subjective naturalness scores even when the underlying model architecture is held constant. source: §5.2, Table 4 evidence: GMVAE-Tacotron trained on LibriTTS at 24 kHz outperforms the 16 kHz LibriTTS models by an average of 0.175 MOS for female speakers and 0.133 MOS for male speakers. confidence: high relevance: high - claim_id: libritts_segmentation_effect_confounded role: complicates claim: Sentence-boundary segmentation and preserved punctuation are theoretically beneficial for TTS training, but their effect is difficult to isolate empirically because corpus size and acoustic quality covary when comparing corpus versions. source: §5.2 evidence: Comparing LibriSpeech (16 kHz, 460h) with LibriTTS (16 kHz, 245h) at matched sampling rate shows a 0.15 MOS gain for female speakers but no consistent gain for male speakers, and the authors explicitly acknowledge in §5.2 that the corpora differ in size, confounding attribution to text-processing quality. confidence: medium relevance: high - claim_id: libritts_snr_filtering_tradeoff role: supports claim: 'SNR-based filtering of audiobook recordings improves corpus quality at the cost of significantly reducing yield: approximately 25% of clean-subset sentences are removed by a 20 dB threshold.' source: §4, Table 3 evidence: A WADA-SNR threshold of 20 dB on the clean subsets discards roughly a quarter of candidate sentences during corpus construction. confidence: high relevance: low - claim_id: libritts_gender_disparity_in_mos role: supports claim: Gender disparities in TTS naturalness persist under multi-speaker training on large open corpora, with male voices showing larger gaps to natural speech than female voices. source: §5.2, Table 4 evidence: The naturalness gap to natural speech is 0.61 MOS for male speakers versus 0.16 MOS for female speakers in the LibriTTS 24 kHz evaluation. confidence: high relevance: high limitations: - Experimental validation uses a single model architecture (GMVAE-Tacotron) and evaluates only six held-out speakers, limiting the generality of the corpus-quality findings. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2010.05646' published_date: "2020-10-12" entry_date: '2026-07-27' year: 2020 venue: arXiv task: - TTS architecture: - GAN relevance: medium evidence_role: - core_evidence current_role: influential method_family: - gan_adversarial_generation claims: - claim_id: hifigan_periodic_structure_quality role: supports claim: Explicitly modeling periodic structure in speech at multiple timescales is necessary for GAN-based vocoders to match the perceptual quality of autoregressive models. source: §2.3, Table 2 evidence: The multi-period discriminator reshapes waveforms into 2D grids indexed by period to model periodicity explicitly; removing it drops MOS from 4.10 to 2.28 in the ablation study. confidence: high relevance: high - claim_id: hifigan_compact_realtime_quality role: supports claim: A compact GAN vocoder can achieve CPU real-time synthesis with quality comparable to autoregressive models when discriminator design, rather than generator depth, is the primary quality bottleneck. source: §4.1, Table 1 evidence: V3 (1.46M parameters) achieves MOS 4.05 at 13.4x faster than real-time on CPU, comparable to WaveNet MoL at a fraction of the compute cost. confidence: high relevance: high - claim_id: hifigan_speaker_generalization_mel_only role: supports claim: Vocoders trained on single-speaker data generalise to unseen speakers when the generator is conditioned only on mel-spectrograms, with quality exceeding flow-based and autoregressive alternatives. source: §4.3, Table 3 evidence: On VCTK's nine held-out speakers, all three HiFi-GAN variants outperform WaveNet MoL, WaveGlow, and MelGAN despite training only on single-speaker mel-spectrogram conditioning, indicating the mel-spectrogram representation itself generalises across speaker identity without explicit speaker conditioning. confidence: medium relevance: low - claim_id: hifigan_finetuning_predicted_spectrograms role: supports claim: Fine-tuning a mel-spectrogram vocoder on predicted (rather than ground-truth) spectrograms substantially improves end-to-end TTS quality, while flow-based vocoders do not benefit from the same adaptation. source: §4.4, Table 4 evidence: In the end-to-end experiment with Tacotron2-predicted mel-spectrograms, fine-tuned HiFi-GAN variants all exceed MOS 4.0, while fine-tuned WaveGlow does not improve over its pre-fine-tuning baseline. confidence: high relevance: high - claim_id: hifigan_discriminator_over_generator role: supports claim: Discriminator architecture choices have a larger impact on GAN vocoder quality than generator architecture choices. source: §4.2, Table 2 evidence: The ablation shows removing MPD causes a catastrophic MOS drop from 4.10 to 2.28, a far larger effect than varying generator capacity across the V1/V2/V3 variants. confidence: high relevance: high limitations: - Evaluation is limited to LJSpeech (single speaker) and VCTK (multi-speaker English); performance on noisy, spontaneous, or cross-lingual speech is not assessed. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2105.06337' published_date: "2021-05-13" entry_date: '2026-07-27' year: 2021 venue: arXiv task: - TTS architecture: - diffusion - transformer-enc-dec relevance: medium evidence_role: - core_evidence current_role: influential method_family: - transformer_encoder_decoder - diffusion_generation claims: - claim_id: gradtts_diffusion_matches_ar_with_speed_control role: supports claim: Diffusion-based acoustic feature generators can match autoregressive TTS naturalness while enabling flexible inference-time speed-quality trade-offs not available in single-pass models. source: §4.1, Table 2 evidence: Grad-TTS-10 reaches MOS 4.38 versus 4.32 for Tacotron 2 while running roughly twice as fast (RTF 0.033 vs. 0.075 on GPU). confidence: high relevance: high - claim_id: gradtts_text_conditioned_prior_reduces_steps role: supports claim: Initialising reverse diffusion from a text-conditioned noise distribution rather than standard Gaussian substantially reduces the number of steps required for high-quality synthesis. source: §3.1, Table 1 evidence: A preference test (p < 0.005) shows that even at 50 reverse-diffusion steps, the standard N(0,I) variant is significantly worse than Grad-TTS-10 using the text-conditioned N(mu,I) prior. confidence: high relevance: high - claim_id: gradtts_higher_likelihood_than_flow role: supports claim: Diffusion models can achieve higher data log-likelihood on mel-spectrograms than normalising-flow models with larger decoder capacity explicitly trained for maximum likelihood. source: §4.2, Table 2 evidence: Grad-TTS achieves a log-likelihood of 0.174 versus 0.082 for Glow-TTS despite Glow-TTS having a 3x larger decoder trained explicitly to maximise likelihood. confidence: medium relevance: low - claim_id: gradtts_diminishing_returns_with_steps role: supports claim: Subjective quality in diffusion-based mel-spectrogram synthesis improves diminishingly with reverse-diffusion step count, with most quality gain recovered by 10 steps rather than 1000. source: §4.1, Table 2 evidence: Grad-TTS-1000 reaches MOS 4.44 versus 4.38 for Grad-TTS-10, a gain of only 0.06 for a 100x increase in reverse-diffusion steps. confidence: high relevance: high limitations: - Evaluation is restricted to a single-speaker English dataset (LJSpeech); no multi-speaker, zero-shot, or multilingual capability is demonstrated. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2204.02152' published_date: "2022-04-05" entry_date: '2026-07-27' year: 2022 venue: arXiv task: - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: influential method_family: [] claims: - claim_id: utmos_listener_dependent_modeling_improves_accuracy role: supports claim: Listener-dependent modeling substantially improves MOS prediction accuracy, particularly for out-of-domain data with limited labeled samples. source: §3.1.3, Table 2b, Table 3b evidence: Removing listener IDs drops system-level SRCC from 0.972 to 0.931 on the OOD track and degrades utterance-level MSE substantially on both tracks. confidence: high relevance: high - claim_id: utmos_contrastive_loss_improves_spearman role: supports claim: Contrastive pairwise ranking losses improve Spearman correlation in MOS prediction more than standard regression losses alone. source: §3.1.2, §4.3, Table 2 evidence: The contrastive ranking loss produces a larger gain on the OOD track than the main track, particularly for ranking (SRCC) metrics. confidence: high relevance: high - claim_id: utmos_ensemble_stacking_more_robust role: supports claim: SSL-based features from multiple heterogeneous pretrained models (wav2vec 2.0, HuBERT, WavLM) combined via ensemble stacking yield more robust MOS predictions than any single model. source: §3.3, §4.4, Tables 4-5 evidence: The full 17-strong + 48-weak learner ensemble achieves utterance-level MSE of 0.165 versus 0.216 for a single strong learner alone. confidence: high relevance: high - claim_id: utmos_phoneme_input_helps_low_resource_not_indomain role: refines claim: Phoneme-level linguistic inputs improve MOS prediction robustly in low-resource conditions but may not benefit in-domain prediction when sufficient labeled data is available. source: §3.1.4, §4.3, Table 2 evidence: 'The phoneme encoder ablation shows mixed results: slightly degrading main-track metrics while improving OOD-track metrics.' confidence: medium relevance: high limitations: - The model is trained and evaluated exclusively on synthetic speech from Blizzard and Voice Conversion Challenges, so its generalisation to modern neural TTS systems (LLM-based, flow-matching) with different failure modes is not characterised in this paper. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2206.04658' published_date: "2022-06-09" entry_date: '2026-07-27' year: 2022 venue: arXiv task: - TTS architecture: - GAN relevance: high evidence_role: - core_evidence current_role: influential method_family: - gan_adversarial_generation claims: - claim_id: bigvgan_periodic_activation_improves_ood_robustness role: supports claim: Periodic activation functions with learnable frequency parameters improve GAN vocoder robustness to out-of-distribution audio, including unseen recording environments, languages, and non-speech signals. source: §3.2, §4.5, §4.6 evidence: On out-of-distribution real-world Korean speech, BigVGAN reaches SMOS 4.38 versus 4.21 for HiFi-GAN. confidence: high relevance: high - claim_id: bigvgan_antialiasing_reduces_artifacts role: supports claim: Anti-aliasing via low-pass filtering around nonlinear activations reduces high-frequency artifacts in neural vocoders and is complementary to periodic inductive bias. source: §3.3, §4.7, Table 4 evidence: Ablation shows removing the low-pass filter degrades OOD performance, and removing both filter and Snake activation degrades it further still, with differences statistically significant (p<0.01). confidence: high relevance: low - claim_id: bigvgan_scaling_benefits_with_stabilization role: supports claim: GAN vocoders benefit significantly from scaling model capacity when paired with appropriate architectural improvements and training stabilisation techniques, whereas naive scaling without those changes leads to early collapse. source: §3.4, §4.7 evidence: Scaling from 14M to 112M parameters requires gradient clipping at global norm 10^3 and a halved learning rate (1e-4) to avoid early training collapse caused by amplified discriminator gradients. confidence: medium relevance: low - claim_id: bigvgan_smos_more_sensitive_than_mos role: refines claim: Similarity MOS (SMOS), which requires direct comparison with ground truth, is a more sensitive evaluation metric than standard MOS for distinguishing speaker fidelity under diverse and out-of-distribution conditions. source: §4.3 evidence: BigVGAN-base (14M params) already outperforms HiFi-GAN on SMOS (+0.05) even though the MOS gap between these small models is comparatively less diagnostic, and the SMOS gap over HiFi-GAN widens further under noisy and OOD conditions (e.g. +0.13 SMOS on real-world Korean). confidence: high relevance: high - claim_id: bigvgan_diverse_training_data_necessary role: supports claim: Training on diverse, multi-condition speech data (including noisy recordings) is a necessary complement to architectural improvements for achieving universal vocoding performance. source: §4.7, Table 5 evidence: All models are trained on the full LibriTTS train set including the noisy train-other-500 subset, which the ablations in Table 5 identify as contributing to OOD robustness. confidence: medium relevance: low limitations: - Trained and evaluated on 24 kHz audio only; generalisation to 44.1 kHz or 48 kHz audio common in music and broadcast is not tested. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2210.13438' published_date: "2022-10-24" entry_date: '2026-07-27' year: 2022 venue: arXiv task: - codec architecture: - GAN relevance: high evidence_role: - core_evidence current_role: influential method_family: - gan_adversarial_generation claims: - claim_id: a_multi_scale_stft_discriminator_alone_is_sufficient role: supports claim: A multi-scale STFT discriminator alone is sufficient for perceptual quality in neural audio codecs, removing the need for waveform-domain discriminators. source: §4.5.1, Table 2 evidence: A key training contribution is the multi-scale STFT (MS-STFT) discriminator, replacing the MSD+Mono-STFT combination from SoundStream. confidence: high relevance: high - claim_id: gradient_balancers_that_normalise_loss_contributions_by_expected role: supports claim: Gradient balancers that normalise loss contributions by expected gradient magnitude substantially stabilise training when combining reconstruction, adversarial, and commitment losses with widely varying natural scales. source: §3.4, Table A.4 evidence: The training objective combines L1 time-domain loss, multi-scale mel-spectrogram loss, adversarial loss, feature matching loss, and RVQ commitment loss. confidence: high relevance: low - claim_id: residual_vector_quantization_supports_variable_bitrate_operation_from role: supports claim: Residual vector quantization supports variable-bitrate operation from a single model by varying the number of active codebooks at inference, with each additional codebook yielding diminishing quality returns. source: §3.2, Table 1 evidence: The RVQ layer uses up to 32 codebooks of 1024 entries each (10 bits per codebook), supporting multiple target bandwidths (1.5, 3, 6, 12, 24 kbps at 24 kHz) by selecting a variable number of codebooks during training. confidence: high relevance: low - claim_id: auxiliary_transformer_language_models_over_rvq_codes_can role: supports claim: Auxiliary Transformer language models over RVQ codes can reduce effective bitrate by 25-40% through entropy coding without perceptual quality degradation, at the cost of increased latency. source: §3.3, §4.5 evidence: Optionally, a 5-layer Transformer language model trained over RVQ codes enables arithmetic entropy coding, reducing effective bitrate by 25-40% without perceptual quality degradation. confidence: high relevance: high - claim_id: neural_audio_codecs_outperform_traditional_dsp_codecs_at role: supports claim: Neural audio codecs outperform traditional DSP codecs at low bitrates across both speech and music domains, with the quality gap widening as bitrate decreases. source: §4.5, Table 1, Figure 3 evidence: For stereophonic 48 kHz music (Table 4), EnCodec at 6 kbps matches MP3 at 64 kbps (both ~82.9 MUSHRA), achieving comparable quality at 10x lower bitrate. confidence: high relevance: high limitations: - The 48 kHz model in non-streamable configuration operates slower than real time on CPU, limiting deployment without GPU acceleration or hardware-specific optimisation. - Arithmetic coding precision issues (floating-point non-determinism across architectures) required a probability rounding workaround that the authors note may be insufficient for practical deployment. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2301.11325' published_date: "2023-01-26" entry_date: '2026-07-27' year: 2023 venue: arXiv task: [] architecture: - autoregressive-LM relevance: high evidence_role: - core_evidence current_role: active_evidence method_family: - autoregressive_generation claims: - claim_id: hierarchical_autoregressive_modeling_over_semantic_and_acoustic_tokens role: supports claim: Hierarchical autoregressive modeling over semantic and acoustic tokens enables long-form music generation (several minutes) with temporal coherence at 24 kHz. source: §3.2, §6 evidence: An ablation removing the semantic modeling stage shows that generating acoustic tokens directly from MuLan tokens preserves FAD scores but degrades KLD (1.01 to 1.05) and MCC (0.51 to 0.49), confirming that the semantic intermediary improves text adherence and long-term structure. confidence: high relevance: low - claim_id: a_joint_audio_text_embedding_space_can_substitute role: supports claim: A joint audio-text embedding space can substitute for paired text-audio supervision at training time, allowing generative models to be trained on audio-only corpora and conditioned on text at inference. source: §3.1, §4.2 evidence: 'The system uses three independently pretrained and frozen models: SoundStream (a neural audio codec operating at 24 kHz, 6 kbps, with 12-level RVQ yielding 600 acoustic tokens per second), w2v-BERT (a 600M-parameter self-supervised model whose intermediate representations are discretized into 25 semantic tokens per second via k-means clustering), and MuLan (a contrastive music-text joint embedding model trained on weakly paired music-text data).' confidence: high relevance: low - claim_id: semantic_token_intermediaries_improve_adherence_to_text_descriptions role: supports claim: Semantic token intermediaries improve adherence to text descriptions in hierarchical audio generation beyond what direct acoustic token prediction achieves. source: §5, Table 1 evidence: An ablation removing the semantic modeling stage shows that generating acoustic tokens directly from MuLan tokens preserves FAD scores but degrades KLD (1.01 to 1.05) and MCC (0.51 to 0.49), confirming that the semantic intermediary improves text adherence and long-term structure. confidence: high relevance: low - claim_id: for_text_conditioned_music_generation_perceptual_audio_quality role: supports claim: For text-conditioned music generation, perceptual audio quality (FAD) and semantic text alignment (MCC, KLD) are complementary evaluation axes that do not always correlate with each other. source: §4.4, Table 1 evidence: 'The system uses three independently pretrained and frozen models: SoundStream (a neural audio codec operating at 24 kHz, 6 kbps, with 12-level RVQ yielding 600 acoustic tokens per second), w2v-BERT (a 600M-parameter self-supervised model whose intermediate representations are discretized into 25 semantic tokens per second via k-means clustering), and MuLan (a contrastive music-text joint embedding model trained on weakly paired music-text data).' confidence: high relevance: high - claim_id: large_autoregressive_audio_lms_trained_on_extensive_unlabeled role: supports claim: Large autoregressive audio LMs trained on extensive unlabeled corpora memorize only a small fraction of training sequences exactly, but approximate semantic matches affect a higher proportion of generated outputs under targeted prompting. source: §5, Figure 3 evidence: Memorization analysis shows exact token matches below 0.2% even when prompting with 10 seconds of training audio, and approximate matches (based on optimal transport between token histograms) below 1%, with detected approximate matches largely corresponding to low-entropy, repetitive token sequences. confidence: high relevance: low limitations: - MCC, one of the two primary text-adherence metrics, is computed using MuLan itself, the same model used for conditioning. - This circularity biases the metric in MusicLM's favour relative to baselines that do not use MuLan representations. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2305.07243' published_date: "2023-05-12" entry_date: '2026-07-27' year: 2023 venue: arXiv task: - TTS architecture: - autoregressive-LM - diffusion - VAE relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: - autoregressive_generation - diffusion_generation - variational_latent_generation claims: - claim_id: conditioning_a_diffusion_decoder_on_the_continuous_latent role: supports claim: Conditioning a diffusion decoder on the continuous latent activations of an autoregressive model rather than its discrete token outputs substantially improves output quality in a cascaded AR-diffusion TTS pipeline. source: §2.2.2, Appendix B.4 evidence: 'The "TorToise Trick" is the most consequential design choice: after training the diffusion decoder on discrete VQVAE tokens to convergence, it is fine-tuned to decode the final hidden-state activations of the autoregressive model rather than the discrete tokens.' confidence: high relevance: high - claim_id: contrastive_re_ranking_of_multiple_autoregressive_candidates_using role: supports claim: Contrastive re-ranking of multiple autoregressive candidates using a text-speech discriminator measurably improves the final output quality of a TTS system without requiring the expensive decoder to process every candidate. source: §2.3, §4 evidence: Third, CLVP (Contrastive Language-Voice Pretrained Transformer) is a CLIP-style dual encoder trained contrastively on text/speech pairs; it scores autoregressive output candidates during re-ranking before the expensive diffusion step, allowing many candidates to be filtered cheaply. confidence: high relevance: low - claim_id: applying_image_generation_scaling_techniques_large_scale_self role: supports claim: Applying image-generation scaling techniques (large-scale self-supervised data, generalist transformer architectures, multi-stage AR-then-diffusion generation) to speech synthesis yields high-expressiveness multi-speaker TTS even when trained by a single researcher on commodity hardware. source: §7 evidence: Second, a GPT-2-style autoregressive decoder (30 transformer layers, dim 1024, 16 heads) is trained to predict sequences of these speech tokens conditioned on text and a speaker conditioning vector. confidence: high relevance: low - claim_id: building_a_large_scale_tts_training_corpus_by role: supports claim: Building a large-scale TTS training corpus by scraping and filtering internet audio (audiobooks, podcasts) with automatic transcription is a viable path to tens-of-thousands-of-hours datasets without manual labelling. source: §5, Appendix A evidence: The extended dataset of 49,000 hours was self-assembled from audiobooks and podcasts, transcribed with a fine-tuned wav2vec2-large model that was adapted to predict punctuation. confidence: high relevance: low limitations: - No formal listening test or MOS table is reported. - The primary quality claim rests on informal sample comparisons; the paper's own evaluation suite (CLVP-FID) is not a standard benchmark. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2306.00814' published_date: "2023-06-01" entry_date: '2026-07-27' year: 2023 venue: arXiv task: - TTS architecture: - GAN relevance: high evidence_role: - core_evidence current_role: active_evidence method_family: - gan_adversarial_generation claims: - claim_id: maintaining_constant_temporal_resolution_throughout_a_gan_vocoder role: supports claim: Maintaining constant temporal resolution throughout a GAN vocoder, with ISTFT as the sole upsampling step, eliminates aliasing artefacts and dramatically reduces inference cost without sacrificing perceptual quality. source: §3.1, §4.3, Table 6 evidence: 'Vocos is a GAN vocoder with an isotropic architecture: the temporal resolution of the feature sequence remains constant throughout the network, matching the hop-size resolution of the STFT.' confidence: high relevance: high - claim_id: implicit_phase_wrapping_via_a_unit_circle_activation role: supports claim: Implicit phase wrapping via a unit-circle activation is essential for stable GAN training of complex-valued spectrogram generators; alternatives that clamp or clip phase angles substantially degrade output quality. source: §3.2, §4.1.1, Table 1 evidence: 'Phase is handled by a unit-circle activation: the raw phase output p is mapped via cosine and sine to obtain a complex coefficient STFT = M * (cos(p) + j*sin(p)), implicitly enforcing phase wrapping into (-pi, pi] without requiring a clipped nonlinearity.' confidence: high relevance: low - claim_id: fourier_domain_vocoders_reduce_periodicity_errors_more_effectively role: supports claim: Fourier-domain vocoders reduce periodicity errors more effectively than time-domain GANs, suggesting that modelling harmonics in the frequency domain provides a stronger inductive bias for voiced speech. source: §4.1.1, Table 1 evidence: In objective metrics, Vocos leads all baselines on VISQOL (4.66) and PESQ (3.70), and achieves the best periodicity score (0.101), indicating fewer harmonic artefacts than time-domain GANs (Table 1). confidence: high relevance: low - claim_id: convnext_blocks_with_isotropic_architecture_outperform_dilated_resblocks role: supports claim: ConvNeXt blocks with isotropic architecture outperform dilated ResBlocks in the Fourier-domain vocoder setting, even though dilated convolutions were motivated by the need to expand receptive fields in time-domain models. source: §4.1.1, Table 1 evidence: Ablations show that the unit-circle phase activation is essential (removing it reduces PESQ from 3.70 to 3.57), ConvNeXt blocks outperform standard ResBlocks (PESQ 3.70 vs 3.53), and Snake activations (used in BigVGAN) do not help in the Fourier-domain setting (Table 1). confidence: high relevance: low - claim_id: a_fourier_domain_gan_vocoder_trained_as_a role: supports claim: A Fourier-domain GAN vocoder trained as a neural codec decoder can substantially improve perceptual quality over the original codec decoder across all bitrates without architectural changes to the upstream codec. source: §4.2, Table 5 evidence: 'For the EnCodec decoder variant, Vocos substantially outperforms the original EnCodec decoder in perceptual quality: at 12 kbps, Vocos scores MOS 4.00 vs EnCodec''s 3.08; at 1.5 kbps the gap is even wider (2.73 vs 1.09) (Table 5).' confidence: high relevance: high limitations: - Vocos's mel-spectrogram MOS scores are reported on LibriTTS using crowd-sourced listeners; the ground-truth MOS (3.81) is noticeably below what might be expected for studio speech, suggesting the evaluation pool or headphone compliance filtering may limit the discriminative power of the subjective test. - Statistical equivalence with BigVGAN is shown, but the test may be underpowered for detecting small differences. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2402.01912' published_date: "2024-02-02" entry_date: '2026-07-27' year: 2024 venue: arXiv task: - TTS architecture: - autoregressive-LM relevance: high evidence_role: - core_evidence current_role: active_evidence method_family: - autoregressive_generation claims: - claim_id: automatic_acoustic_labeling_can_substitute_for_human_annotations role: supports claim: Automatic acoustic labeling can substitute for human annotations in training large-scale instruction-conditioned speech language models without a loss in attribute control accuracy relative to human-labeled systems. source: §3.1, §3.2, §4.1 evidence: Objective attribute control is validated using the same automatic classifiers that labeled the training data. confidence: high relevance: high - claim_id: including_a_small_proportion_of_high_fidelity_audio role: supports claim: Including a small proportion of high-fidelity audio (approximately 1%) in a predominantly noisy training corpus, combined with explicit recording-quality labels, enables a speech LM to generate professional-sounding speech on demand from text prompts alone. source: §3.1.2, §4.2, Table 1 evidence: 'High audio fidelity is achieved through a combination of: (1) using DAC rather than EnCodec, which the authors report provides subjectively and objectively superior audio quality, and (2) including LibriTTS-R in training, which provides approximately 500 hours (roughly 1%) of clean, professionally enhanced speech alongside the crowd-sourced data.' confidence: high relevance: high - claim_id: the_choice_of_neural_audio_codec_has_a role: supports claim: The choice of neural audio codec has a measurable effect on perceptual audio quality in autoregressive TTS; higher-fidelity codecs translate directly to higher MOS and objective quality scores. source: §3.3, §4.2, Table 1–2 evidence: On samples prompted for excellent recording quality, the higher-fidelity codec system outperforms Audiobox on PESQ, STOI, and SI-SDR and approaches the ground-truth values. confidence: high relevance: high - claim_id: natural_language_conditioning_on_accent_can_be_achieved role: supports claim: Natural language conditioning on accent can be achieved in a single TTS model covering dozens of accents, though classifier accuracy reflects the noise and imbalance inherent in automatic accent labeling of crowd-sourced data. source: §3.1.1, §4.1 evidence: A large language model (Stable Beluga 2) then converts keyword tuples into natural language sentences, e.g. "a woman with a deep voice speaking slowly and somewhat monotonously with a Hungarian accent in an echoey room with background noise." This approach is applied to the 45k-hour English Multilingual LibriSpeech corpus and the smaller 585-hour LibriTTS-R corpus (included for its higher audio fidelity from the Miipher speech-enhancement model). confidence: high relevance: high limitations: - The evaluation compares only against Audiobox. - No standard TTS baselines (reference-based zero-shot systems, encoder-decoder models) are included, making it impossible to assess whether the MOS gains arise from the conditioning approach, the codec choice, or the training data mix. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2403.16973' published_date: "2024-03-25" entry_date: '2026-07-27' year: 2024 venue: arXiv task: - TTS architecture: - autoregressive-LM relevance: high evidence_role: - core_evidence current_role: active_evidence method_family: - autoregressive_generation claims: - claim_id: autoregressive_codec_language_models_can_perform_speech_infilling role: supports claim: Autoregressive codec language models can perform speech infilling with naturalness approaching that of the original unedited recording when masked spans are relocated to the end of the sequence, enabling bidirectional context conditioning within a causal framework. source: §3.1, §5.3, Table 5 evidence: 'The core mechanism is a two-step token rearrangement procedure applied before autoregressive modeling: In the first step (causal masking), the spans to be generated are relocated from their original position in the codec matrix to the end of the sequence.' confidence: high relevance: high - claim_id: zero_shot_tts_and_speech_editing_can_be role: supports claim: Zero-shot TTS and speech editing can be unified as a single autoregressive infilling operation without task-specific architectural components, at no cost to performance on either task. source: §3.4, §5.4, Table 6 evidence: VoiceCraft reframes both speech editing (infilling) and zero-shot TTS (continuation) as left-to-right language modeling over rearranged neural codec token sequences. confidence: high relevance: low - claim_id: wer_measured_by_asr_systems_is_an_unreliable role: supports claim: 'WER measured by ASR systems is an unreliable proxy for perceptual intelligibility when evaluating speech synthesis quality: systems can achieve lower WER than ground truth recordings while receiving substantially lower intelligibility ratings from human listeners.' source: §5.3, §5.4 evidence: 'A recurrent observation across both tasks: FluentSpeech achieves lower WER than the ground truth (3.5% vs 3.8% for TTS; 4.5% vs 5.4% for editing), yet receives substantially lower intelligibility MOS from human listeners.' confidence: high relevance: high - claim_id: evaluation_of_speech_synthesis_exclusively_on_audiobook_data role: supports claim: Evaluation of speech synthesis exclusively on audiobook data underestimates the performance gap between systems when applied to in-the-wild recordings with diverse accents, noise, and speaking styles. source: §5.3, §5.4 evidence: This suggests that ASR-based WER can be a misleading proxy for perceptual intelligibility when systems produce dull or robotic speech that ASR models handle well. confidence: high relevance: high - claim_id: scaling_autoregressive_codec_lm_parameters_consistently_improves_objective role: supports claim: Scaling autoregressive codec LM parameters consistently improves objective metrics across intelligibility and acoustic fidelity measures, with larger gaps between larger model sizes suggesting further gains from continued scaling. source: §5.2, Table 3 evidence: Ablation over model sizes (120M, 430M, 830M) shows consistent gains from scaling across WER, MCD, F0, and energy distance metrics. confidence: high relevance: low limitations: - The inference-time artifact mitigation (generating 10 candidates and discarding the 4 longest) adds significant latency and compute cost. - The strategy is acknowledged as inelegant, and the underlying cause (repetitive loop generation) is unresolved. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2406.00654' published_date: "2024-06-02" entry_date: '2026-07-27' year: 2024 venue: arXiv task: - TTS architecture: - autoregressive-LM relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: - autoregressive_generation claims: - claim_id: standard_supervised_training_objectives_for_tts_produce_a role: supports claim: Standard supervised training objectives for TTS produce a systematic mismatch with human perceptual evaluation metrics such as MOS and WER, and correcting this mismatch through preference-aware fine-tuning yields large performance gains. source: §1, §4.2, Table 1 evidence: The framework is applied as a fine-tuning pass on a pre-trained VoiceCraft model (330M or 830M parameters), which uses SpeechTokenizer (EnCodec-based, 4 RVQ codebooks, vocabulary of 2048) for discrete acoustic tokenisation. confidence: high relevance: high - claim_id: existing_rlhf_methods_requiring_pairwise_preference_data_from role: supports claim: Existing RLHF methods requiring pairwise preference data from the same input (DPO) are difficult to apply directly to autoregressive codec TTS because these models lack sufficient output diversity to form meaningful preference pairs from a fixed transcript-prompt combination. source: §4.2, Appendix B evidence: In the sampling stage, a fixed target transcript is paired with a batch of distinct speech prompts drawn from an unseen speaker pool, yielding a diverse set of synthesised outputs; varying the prompt (rather than stochastic decoding) provides much richer quality variation across samples. confidence: high relevance: high - claim_id: uncertainty_in_human_speech_quality_annotations_is_not role: supports claim: Uncertainty in human speech quality annotations is not noise to be discarded but an informative signal that, when incorporated into the optimization objective, improves the consistency of generated speech across listeners. source: §4.2, §6.3, Table 3 evidence: An ablation without uncertainty weighting (UNO-null) achieves similar mean MOS but higher output variance, confirming that uncertainty modelling specifically improves generation consistency rather than average quality. confidence: high relevance: high - claim_id: rlhf_style_alignment_for_tts_can_be_achieved role: supports claim: RLHF-style alignment for TTS can be achieved with a small number of self-generated samples (hundreds) without access to ground truth speech, making it practical for post-training fine-tuning at low computational cost. source: §4.1, §5, Appendix D evidence: Only 400 self-generated samples (balanced 200 positive / 200 negative) are needed, requiring approximately 10 minutes of fine-tuning on a single A100 GPU. confidence: high relevance: low - claim_id: alignment_objectives_designed_for_naturalness_mos_transfer_to role: supports claim: Alignment objectives designed for naturalness MOS transfer to other perceptual dimensions such as emotion by substituting the selection criterion, demonstrating that preference-based fine-tuning generalises beyond a single quality axis. source: §6.4, Table 4 evidence: 'The method extends naturally to emotion TTS by substituting the selection criterion in the sampling stage: using emotional ESD dataset prompts and sorting by valence or arousal score instead of MOS drives the model to generate speech with target affective properties.' confidence: high relevance: high limitations: - 'The comparison with SpeechAlign is acknowledged by the authors to be partially unfair: SpeechAlign-DPO requires ground truth speech as positive samples during optimization, which is additional supervision not available to UNO.' - Presenting both as baselines without fully separating this distinction may understate SpeechAlign's performance under matched conditions. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2406.18009' published_date: "2024-06-26" entry_date: '2026-07-27' year: 2024 venue: arXiv task: - TTS architecture: - flow-matching relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - flow_matching_generation claims: - claim_id: explicit_phoneme_alignment_supervision_in_non_autoregressive_tts role: supports claim: Explicit phoneme alignment supervision in non-autoregressive TTS, while useful for early convergence, imposes a naturalness ceiling that joint character-level training can exceed. source: §3.4, Table 2 evidence: 4) shows E2 TTS takes longer to converge than Voicebox (which benefits from explicit alignment supervision early in training) but surpasses it by end of training, consistent with the hypothesis that joint learning of duration and acoustic modeling yields a better end state. confidence: high relevance: high - claim_id: flow_matching_mel_spectrogram_generators_can_learn_alignment role: supports claim: Flow-matching mel spectrogram generators can learn alignment implicitly from raw characters and filler tokens, eliminating the need for grapheme-to-phoneme converters, phoneme aligners, and duration models. source: §2.1, §2.3 evidence: 'The CMOS comparison directly identifies phoneme alignment as the primary bottleneck to naturalness: replacing it with joint character-and-filler modeling closes the gap.' confidence: high relevance: high - claim_id: unsupervised_pretraining_on_unlabeled_speech_improves_downstream_zero role: supports claim: Unsupervised pretraining on unlabeled speech improves downstream zero-shot TTS performance in both intelligibility and speaker similarity. source: §3.4, Table 1 evidence: On LibriSpeech-PC test-clean (1,132 samples, 39 speakers), E2 TTS trained on Libriheavy achieves WER 2.0% and speaker similarity (SIM-o) 0.675 from random initialization, improving to WER 1.9% and SIM-o 0.708 with unsupervised pretraining. confidence: high relevance: high - claim_id: zero_shot_tts_speaker_similarity_scores_at_inference role: supports claim: Zero-shot TTS speaker similarity scores at inference can exceed those of ground-truth recordings on standard speaker verification metrics, suggesting the metrics reward consistency within a generation rather than perceptual identity. source: §3.4, Table 2 evidence: On LibriSpeech-PC test-clean (1,132 samples, 39 speakers), E2 TTS trained on Libriheavy achieves WER 2.0% and speaker similarity (SIM-o) 0.675 from random initialization, improving to WER 1.9% and SIM-o 0.708 with unsupervised pretraining. confidence: high relevance: high - claim_id: non_autoregressive_tts_systems_trained_jointly_on_duration role: supports claim: Non-autoregressive TTS systems trained jointly on duration and acoustics scale predictably with training data volume without requiring architectural changes. source: §3.4, Table 1 evidence: Training on 200,000 hours of proprietary data reaches WER 1.9% and SIM-o 0.707 from random initialization. confidence: high relevance: low limitations: - The CMOS evaluation covers only 39 samples from 39 LibriSpeech speakers read English. - 'The naturalness finding ("indistinguishable from ground truth") is thus narrow in scope: spontaneous speech, non-native speakers, expressive or prosodically complex content, and cross-lingual settings are untested.' caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2407.05361' published_date: "2024-07-07" entry_date: '2026-07-27' year: 2024 venue: arXiv task: - TTS - evaluation architecture: [] relevance: medium evidence_role: - core_evidence - infrastructure current_role: active_evidence method_family: [] claims: - claim_id: emilia_itw_data_matches_formal_improves_spontaneous role: supports claim: Training speech generation models on large-scale in-the-wild data, after automated quality filtering, can match audiobook-derived training data on formal-domain benchmarks while improving performance on spontaneous speech generation. source: §V.A, Table V evidence: AR+SoundStorm and VoiceBox trained on Emilia-En (46k hours) match MLS-trained counterparts on LibriSpeech-Test WER/S-SIM/FSD, but improve WER, S-SIM, and FSD on Emilia-Test (spontaneous speech) for both architectures. confidence: high relevance: low - claim_id: emilia_scaling_law_diminishing_returns role: supports claim: Increasing in-the-wild training-set size continues to improve zero-shot TTS quality at fixed model capacity, with the largest gains at smaller scales and diminishing but still-positive returns as scale grows. source: §V.B, Figure 5 evidence: Scaling training data from 5k to 134k hours steadily reduces AR+SoundStorm WER (5.2% to 4.2% on LibriSpeech-Test; 5.7% to 4.9% on Emilia-Test), with gains concentrated below 46k hours and convergence emerging around 100k hours. confidence: high relevance: low - claim_id: emilia_multilingual_pooling_tradeoff role: complicates claim: Training a single model on pooled multilingual in-the-wild data enables crosslingual generation but trades off same-language quality relative to a monolingual model trained on comparable per-language data volume. source: §V.C evidence: The Emilia-Large-trained multilingual AR+SoundStorm model scores WER 4.9%, S-SIM 0.636, and FSD 15.31 on English, all slightly worse than an English-only 134k-hour model (WER 4.5%, S-SIM 0.65, FSD 14.8). confidence: medium relevance: low - claim_id: emilia_filtering_discards_majority_of_raw_audio role: complicates claim: Automated quality-filtering pipelines for in-the-wild speech data necessarily discard a large majority of raw source audio to reach a training-quality bar. source: §III, Table 2 evidence: On a 666.94-hour benchmark sample, Emilia-Pipe's six stages retain only 38.75% of the input (258.44 hours, 96,117 clips) after DNSMOS, language-confidence, and duration-outlier filtering. confidence: high relevance: high - claim_id: emilia_crosslingual_transfer_remains_harder role: complicates claim: Crosslingual voice transfer in zero-shot TTS remains substantially harder than same-language transfer even when the training corpus itself is multilingual and balanced across the target languages. source: §V.C, Table VI evidence: Crosslingual WER and speaker similarity degrade relative to matched-language pairs across all six languages, e.g. VoiceBox S-SIM falls from 0.588 En-En to 0.386 En-Zh. confidence: high relevance: high limitations: - Subjective evaluation in the audiobook-vs-in-the-wild comparison uses only sixteen samples and twelve English-speaking evaluators, limiting statistical reliability of reported CMOS/SMOS differences. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2407.08551' published_date: "2024-07-11" entry_date: '2026-07-27' year: 2024 venue: arXiv task: - TTS architecture: - autoregressive-LM relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - autoregressive_generation claims: - claim_id: continuous_mel_spectrogram_representations_preserve_more_speaker_relevant role: supports claim: Continuous mel-spectrogram representations preserve more speaker-relevant acoustic information than vector-quantized codec codes at standard compression rates. source: §5.1, Table 1 evidence: This replaces the top-p discrete sampling of codec language models with a learned continuous sampling mechanism, providing output diversity without requiring manual sampling configuration. confidence: high relevance: low - claim_id: autoregressive_tts_models_trained_to_predict_continuous_frames role: supports claim: Autoregressive TTS models trained to predict continuous frames can achieve naturalness comparable to human speech while avoiding the silence and repetition failures endemic to discrete codec language models. source: §5.2, Table 3 evidence: This replaces the top-p discrete sampling of codec language models with a learned continuous sampling mechanism, providing output diversity without requiring manual sampling configuration. confidence: high relevance: high - claim_id: variational_sampling_in_the_continuous_latent_space_is role: supports claim: Variational sampling in the continuous latent space is more effective than top-p discrete sampling for improving output diversity and speaker similarity in autoregressive TTS. source: §5.3, Table 4 evidence: This replaces the top-p discrete sampling of codec language models with a learned continuous sampling mechanism, providing output diversity without requiring manual sampling configuration. confidence: high relevance: high - claim_id: a_reduction_factor_that_predicts_multiple_frames_per role: supports claim: A reduction factor that predicts multiple frames per autoregressive step can substantially reduce inference time with only modest degradation in speaker similarity. source: §5.4, Table 5 evidence: 'The reduction factor r provides a throughput knob: r=2 halves AR steps and inference time with modest speaker similarity loss, while r=4 reduces inference time to one quarter of the baseline and still outperforms most VALL-E variants on WER.' confidence: high relevance: high limitations: - The subjective evaluation rests on only 40 samples from a single English corpus (LibriSpeech test-clean). - The naturalness and speaker similarity advantages may not generalize to noisier prompts, non-native accents, or other languages. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2409.05377' published_date: "2024-09-09" entry_date: '2026-07-27' year: 2024 venue: arXiv task: - codec architecture: - GAN - VAE relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - gan_adversarial_generation - variational_latent_generation claims: - claim_id: at_very_low_bitrates_around_1_kbps_model role: supports claim: At very low bitrates (around 1 kbps), model capacity is a decisive factor in reconstruction quality, with larger models substantially outperforming architecturally sophisticated but smaller codecs. source: §III-C, §IV-B, Table I evidence: The inference RTF of BigCodec on CPU is 1.1x, just barely real-time, compared to 3.1x for BigCodec-base, reflecting the compute cost of the larger model. confidence: high relevance: low - claim_id: low_dimensional_vector_quantisation_before_codebook_lookup_substantially role: supports claim: Low-dimensional vector quantisation before codebook lookup substantially improves codebook utilisation in single-codebook, single-quantisation-step codec designs. source: §III-A, §IV-B evidence: 'Third, low-dimensional vector quantization: latent variables are projected into an 8-dimensional space before quantization against a single codebook of 8192 codes, avoiding the sparse high-dimensional quantization problem that leads to poor codebook utilisation.' confidence: high relevance: low - claim_id: adding_sequential_lstm_modelling_to_a_convolutional_codec role: supports claim: Adding sequential (LSTM) modelling to a convolutional codec encoder improves both perceptual quality and speaker similarity at low bitrates, independently of the parameter count effect. source: §III-A, §IV-D, Table III evidence: The LSTM contributes independently of the parameter count increase it brings. confidence: high relevance: high - claim_id: scaling_codec_model_size_beyond_a_saturation_point role: supports claim: Scaling codec model size beyond a saturation point (approximately 159M parameters in this setting) yields no further reconstruction benefit, analogous to scale-up behaviour observed in neural vocoders. source: §IV-D, Table III evidence: Scaling beyond 159M (BigCodec-300M) and scaling training data to 60k hours (LibriLight) both yield no measurable improvement, indicating a capacity saturation point. confidence: high relevance: low - claim_id: increasing_training_data_volume_from_960_hours_to role: supports claim: Increasing training data volume from 960 hours to 60k hours does not improve codec reconstruction quality, suggesting that model capacity rather than data quantity is the binding constraint at this bitrate. source: §IV-D, Table III evidence: Scaling beyond 159M (BigCodec-300M) and scaling training data to 60k hours (LibriLight) both yield no measurable improvement, indicating a capacity saturation point. confidence: high relevance: low limitations: - BigCodec is trained exclusively on clean English speech (LibriSpeech 960h), while competing codecs such as EnCodec and DAC train on diverse multilingual datasets including music and environmental sounds. - The multilingual generalisation result is encouraging, but the clean-speech-only training domain limits applicability to noisy or music-heavy audio without fine-tuning. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2410.17196' published_date: "2024-10-22" entry_date: '2026-07-27' year: 2024 venue: arXiv task: - evaluation - SCA architecture: [] relevance: high evidence_role: - core_evidence current_role: active_evidence method_family: [] claims: - claim_id: voicebench_pipeline_beats_e2e role: supports claim: Pipeline-based voice assistants (ASR plus LLM) substantially outperform end-to-end audio LLMs on spoken instruction understanding, with the gap persisting even for state-of-the-art proprietary end-to-end systems. source: §3.3, Table 3 evidence: The naive Whisper + LLaMA-3.1-8B pipeline scores 79.1 on the composite VoiceBench score versus 55.7 for the best open-source end-to-end model (DiVA); even GPT-4o-Audio (86.4) trails the Naive-4o pipeline (87.2). confidence: high relevance: low - claim_id: voicebench_e2e_training_degrades_llm_capability role: supports claim: End-to-end audio LLM training can degrade the underlying LLM's text-instruction capabilities, including safety alignment, compared to using the same base model in a pipeline configuration. source: §3.3 evidence: LLaMA-Omni's safety refusal rate drops from 98.5% (text) to 11.4% (speech), and its text-instruction performance falls over 11 points compared to the same base LLM used in a pipeline. confidence: high relevance: low - claim_id: voicebench_mispronunciation_worse_than_disfluency role: refines claim: Mispronunciation causes significantly greater performance degradation in voice assistants than disfluencies or grammatical errors, because it increases effective word error rate rather than merely adding irrelevant context. source: §6.2, Table 5 evidence: Mispronunciation causes an average 20.3% score reduction across models, versus 12.6% for repair disfluencies, while grammatical errors are largely tolerated. confidence: high relevance: low - claim_id: voicebench_clean_synthetic_speech_overestimates_robustness role: contradicts claim: Evaluations using clean synthetic speech substantially overestimate voice assistant robustness relative to real recorded speech, particularly for models with non-standard speech encoders. source: §3.3, Table 4 evidence: Real accent data (SD-QA) is shown to be substantially harder than synthetic accent data for the same models, particularly for low-resource accents such as Philippine and Indian English. confidence: high relevance: low - claim_id: voicebench_encoder_choice_dominates_robustness role: refines claim: The choice of speech encoder determines a voice assistant's robustness to accent and speaker variation more than the choice of base LLM. source: §4.2 evidence: VITA, which uses a non-Whisper CNN+Transformer speech encoder, shows the greatest accent sensitivity among all evaluated systems, a pattern the authors attribute to encoder choice rather than the base LLM. confidence: medium relevance: low limitations: - Evaluates only the text content of responses, not speech output quality; Moshi, the only tested model that natively outputs speech, is scored under the same text-only protocol as text-output models. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.chipsal-1.18 published_date: "2025-01-19" entry_date: '2026-07-27' year: 2025 venue: CHiPSAL 2025 task: - TTS - evaluation architecture: - GAN - transformer-enc-dec relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: - transformer_encoder_decoder - gan_adversarial_generation claims: - claim_id: nepali_gan_vocoder_lower_latency role: supports claim: GAN-based vocoders offer substantially lower inference latency than autoregressive vocoders when generating speech from mel-spectrograms. source: §5.1, Table 4 evidence: Tacotron2 + MelGAN synthesises speech in 0.142s versus 1,320s for Tacotron2 + WaveNet on an NVIDIA GTX 960M GPU, owing to MelGAN's parallel waveform generation. confidence: high relevance: low - claim_id: nepali_gan_vocoder_comparable_quality role: supports claim: GAN-based vocoders achieve comparable or superior perceptual quality to autoregressive vocoders in low-resource language TTS. source: §5.2, Tables 5-7; §3.4.2, Figure 2 evidence: Tacotron2 + MelGAN outperforms Tacotron2 + WaveNet in MOS on both Nepali test sets (4.245 vs. 3.65 on OpenSLR; 2.885 vs. 2.31 on male voice data), confirmed by lower MCD on the OpenSLR set. confidence: medium relevance: high - claim_id: nepali_multispeaker_data_lowers_naturalness role: complicates claim: Naturalness scores for Nepali TTS systems trained on multi-speaker male voice data are substantially lower than those achieved on curated single-accent female corpora. source: §5.2, Tables 5-7 evidence: Both Tacotron2 + MelGAN and Tacotron2 + WaveNet score more than 1 MOS point lower on the male voice dataset relative to OpenSLR, with the gap persisting across both objective (MCD) and subjective (MOS) metrics. confidence: medium relevance: high - claim_id: nepali_low_resource_lags_english_benchmarks role: complicates claim: Performance of well-studied vocoders on low-resource languages lags behind their English-language benchmarks, even when the acoustic model and training procedure are unchanged. source: §5.3, Table 8 evidence: Tacotron2 + WaveNet achieves 3.65 MOS on Nepali OpenSLR versus 4.53 MOS reported for WaveNet on US English data; Tacotron2 + MelGAN achieves 2.885 MOS on the male voice set, below MelGAN's 3.49-3.61 MOS on English. confidence: medium relevance: high limitations: - MOS evaluation is based on only 40 participants rating four samples per model, too small a listener pool for statistically robust conclusions; no significance tests are reported. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.coling-main.352 published_date: "2025-01-19" entry_date: '2026-07-27' year: 2025 venue: COLING task: - TTS architecture: - diffusion - transformer-enc-dec relevance: high evidence_role: - core_evidence current_role: active_evidence method_family: - transformer_encoder_decoder - diffusion_generation claims: - claim_id: applying_diffusion_based_classifier_free_guidance_to_low role: supports claim: Applying diffusion-based classifier-free guidance to low-dimensional prosodic features (pitch, energy, duration) rather than the full Mel-spectrogram achieves faster inference than Mel-diffusion baselines while retaining prosodic diversity benefits over regression-based models. source: §3.2, Table 1 evidence: On synthesis speed (RTF), DiffStyleTTS (0.048) is faster than Grad-TTS (0.250) and Guided-TTS (0.479) because it runs diffusion only over low-dimensional prosodic features rather than full Mel-spectrograms. confidence: high relevance: low - claim_id: hierarchical_conditioning_combining_coarse_grained_implicit_style_tokens role: supports claim: Hierarchical conditioning combining coarse-grained implicit style tokens with fine-grained explicit prosodic diffusion outperforms both purely regression-based and discriminator-constrained diffusion baselines on naturalness and prosodic distribution accuracy. source: §3.2, Table 1 evidence: Coarse-grained implicit style conditions (ISC) are extracted from reference Mel-spectrograms via a GST module (10 tokens, 256-D token embeddings, 4-head attention). confidence: high relevance: high - claim_id: classifier_free_guidance_scale_controls_a_trade_off role: complicates claim: Classifier-free guidance scale controls a trade-off between prosodic diversity and naturalness quality in diffusion-based prosody prediction, with distortion artefacts emerging at high guiding scales. source: §3.3, Table 2 evidence: A dynamic thresholding improvement corrects the standard deviation of the guided noise output to prevent phoneme distortion (elongation/noise artefacts) that arises at high guiding scales (η ≥ 7.0). confidence: high relevance: high - claim_id: dynamic_thresholding_applied_to_cfg_noise_outputs_effectively role: supports claim: Dynamic thresholding applied to CFG noise outputs effectively mitigates phoneme distortion caused by excessive guiding scales in prosodic diffusion models. source: §2.2, §3.3 evidence: A dynamic thresholding improvement corrects the standard deviation of the guided noise output to prevent phoneme distortion (elongation/noise artefacts) that arises at high guiding scales (η ≥ 7.0). confidence: high relevance: low - claim_id: gst_based_implicit_style_conditioning_enables_meaningful_prosodic role: supports claim: GST-based implicit style conditioning enables meaningful prosodic transfer across speakers, outperforming both regression-based style transfer and discriminator-constrained diffusion in subjective preference evaluations. source: §3.4, Table 3 evidence: 'The conditional diffusion module implements classifier-free guidance (CFG) using two jointly trained denoisers sharing the same bidirectional dilated convolution architecture (12 residual layers, WaveNet-style): one conditioned on both text embeddings and implicit style conditions (Ψ_θ1), one conditioned on text embeddings only (Ψ_θ2).' confidence: high relevance: high limitations: - Speaker timbre and prosody are not disentangled — the model can transfer prosodic style but cannot fully separate it from speaker identity. The GST tokens show partial overlap in t-SNE visualizations, indicating shared prosodic modes and incomplete independence. The evaluation is conducted on a private Mandarin Chinese dataset (9 male speakers from similar genres), limiting generalization claims across languages, genders, or more varied speaking styles. The 200-step diffusion process at inference is still far slower than FastSpeech2, and no acceleration scheme (consistency models, flow-matching) is explored. Whether the CFG approach generalises to cross-lingual or highly emotional speech remains untested. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2409.20007' published_date: "2025-01-27" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - SCA architecture: - transformer-enc-dec relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - transformer_encoder_decoder claims: - claim_id: desta2_broad_instruction_following_without_tuning_data role: supports claim: Speech language models can acquire broad instruction-following capabilities over paralinguistic attributes without task-specific speech instruction-tuning data. source: §V.A, Table I evidence: DeSTA2 achieves 56.78% overall accuracy on Dynamic-SUPERB across 48 tasks, surpassing all prior end-to-end systems that used extensive instruction-tuning, using only LLM-generated speech captions with a single generic prompt. confidence: high relevance: low - claim_id: desta2_constrained_prompts_degrade_general_capability role: complicates claim: Automated speech instruction-tuning pipelines that use constrained prompts or QA formatting may degrade the base LLM's general language capabilities. source: §V.B, Table III evidence: Ablations show training with open-ended QA pairs improves degradation-task performance but causes measurable drops in content and semantic dimensions relative to DeSTA2's unconstrained caption-generation approach. confidence: medium relevance: low - claim_id: desta2_distributional_alignment_reduces_forgetting role: supports claim: Distributional alignment between training targets and the base LLM's pretraining regime reduces catastrophic forgetting in speech-extended language models. source: §V.D, Table V evidence: DeSTA2 preserves chain-of-thought reasoning and output-format following (e.g., computing words-per-second, responding in all-caps on demand), capabilities absent from Qwen2-Audio despite its larger training budget. confidence: medium relevance: low - claim_id: desta2_paralinguistic_metadata_drives_performance role: refines claim: Rich paralinguistic metadata coverage, rather than task diversity, is the primary driver of broad spoken language model performance on multi-dimensional speech benchmarks. source: §IV.A, §V.A, Table I evidence: DeSTA2 uses a single captioning task with 12 paralinguistic attributes and outperforms systems with diverse task-specific instruction sets on the paralinguistic dimension of Dynamic-SUPERB (43.14% vs. 29.21% for Qwen2-Audio). confidence: medium relevance: low limitations: - Speaker verification performance remains below specialized baselines (42.5% vs. 47.9% for Qwen2-Audio on the Dynamic-SUPERB speaker dimension); the approach does not address multi-speaker comparison tasks. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.computel-main.6 published_date: "2025-03-01" entry_date: '2026-07-27' year: 2025 venue: ComputEL task: - TTS - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: quantitative_qualitative_dimensions_diverge role: supports claim: Quantitative accuracy ratings and qualitative community feedback reveal distinct dimensions of TTS quality that do not co-vary for low-resource Indigenous language systems. source: §6, Table 1 evidence: Participants rated phonetic accuracy highly (11 "Strongly Agree" / 8 "Agree" aggregated across tasks) while rating contextual appropriateness and naturalness much lower, and identified natural rhythm and tone modulation as absent from the system. confidence: medium relevance: high - claim_id: small_l1_populations_limit_perceptual_eval role: complicates claim: Small L1 speaker populations make standard large-scale perceptual evaluation of Indigenous language TTS impractical, limiting how generalisable accuracy findings can be. source: §Limitations evidence: Five participants constituted the entire available pool; the paper explicitly flags sample size as a limitation and uses qualitative PD methods as an alternative to quantitative MOS-style evaluation. confidence: high relevance: high - claim_id: adoption_depends_on_more_than_quality role: supports claim: Community adoption of TTS tools for language education depends on factors beyond output quality, including interface discoverability, voice representation, and alignment with cultural pedagogy. source: §7.1, §7.2 evidence: The Ojibwe TTS feature had been available for teacher use but was unknown to participants before this workshop; teacher priorities surfaced in the PD sessions included voice diversity, accessible UI labelling, and offline download capability for rural use. confidence: medium relevance: low - claim_id: participatory_design_as_evaluation_method role: refines claim: Participatory design workshops can serve as an evaluation method for Indigenous language TTS when quantitative listener studies are not feasible, eliciting both usability feedback and pedagogical priorities in a single session. source: §4, §8.2 evidence: The combined pre-workshop guided trial and co-design session produced both Likert-scale quality ratings and lesson-plan artefacts, with unprompted participant insights addressing all four research questions without formal prompting. confidence: medium relevance: high limitations: - The study involves five participants from a single community and a single TTS system; the workshop findings, while rich, cannot be generalised quantitatively and are specific to Border Lakes Ojibwe and the Anishinaabemodaa platform context. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.findings-naacl.298 published_date: "2025-04-29" entry_date: '2026-07-27' year: 2025 venue: NAACL task: - TTS - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: instruction_tts_amplifies_occupational_gender_stereotypes role: supports claim: Instruction-conditioned TTS models reflect and amplify occupational gender stereotypes present in training data when interpreting ambiguous style prompts. source: §4.3, Table 1 evidence: Across four Parler-TTS variants and 109 occupations, chi-square analysis shows statistically significant overrepresentation of male voices for male-coded occupations and female voices for female-coded occupations, with standardized residuals up to 27.2 for individual occupation-model pairs. confidence: high relevance: low - claim_id: prompt_mitigation_causes_compensatory_bias role: complicates claim: Prompt-based fairness interventions can introduce compensatory bias rather than achieving gender neutrality in instruction-conditioned TTS. source: §4.4, Tables 2-4 evidence: Methods 2 and 3 (adding explicit fairness language to style prompts) frequently reverse the bias direction for traditionally masculine occupations across all four models, shifting overrepresentation from male to female rather than reducing bias magnitude. confidence: high relevance: low - claim_id: bias_magnitude_varies_across_model_variants role: complicates claim: Gender bias magnitude in instruction-guided TTS varies substantially across model variants trained on the same data distribution, limiting the generalizability of any single model's bias profile. source: §4.3 evidence: Large v1 and Mini v1 share training data but display different bias patterns by occupation; Mini v0.1 and Mini Expresso (fine-tuned variant) likewise differ, suggesting model scale and fine-tuning interact with bias in ways not fully explained by data alone. confidence: medium relevance: low - claim_id: bias_eval_requires_baseline_control_groups role: supports claim: Evaluating gender bias in TTS requires controlling for the model's inherent gendered tendency in baseline generation, separate from the effect of specific style prompts. source: §3.3, §4.2 evidence: Even with an empty style prompt or neutral non-occupational prompts, Parler-TTS models generate voices skewed toward a particular gender, necessitating three distinct control groups to isolate the effect of occupation-related prompting. confidence: high relevance: low limitations: - The analysis is restricted to Parler-TTS, the only publicly available open-source model family accepting free-form style prompts without reference audio at the time of publication; findings may not generalise to other model families. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.naacl-demo.21 published_date: "2025-04-29" entry_date: '2026-07-27' year: 2025 venue: NAACL task: - SCA architecture: [] relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: [] claims: - claim_id: cascaded_pipelines_higher_output_quality_than_e2e role: supports claim: Cascaded spoken dialogue pipelines built from dedicated, task-specific ASR and TTS components can produce substantially higher audio output quality than current end-to-end spoken dialogue systems that generate speech directly from audio input. source: §7, Table 4 evidence: On Switchboard Eval 2000, the directly-synthesized speech from the E2E system Mini-Omni scores UTMOS 2.88, versus 3.97-4.07 for cascaded pipelines using LJSpeech, LibriTTS, or VCTK VITS as the TTS module. confidence: medium relevance: high - claim_id: e2e_coherent_but_repetitive_across_turns role: complicates claim: End-to-end spoken dialogue systems that respond directly to audio input without explicit dialogue-context modeling can produce coherent individual responses while still exhibiting unnatural conversational behavior such as repeated or templated phrasing across turns. source: §7, Table 3 evidence: The text dialogue evaluation finds that Mini-Omni "produces highly coherent and contextually relevant responses" but "exhibits significant overlap in its responses, often generating template-like outputs," which the authors note could make conversations feel less natural in deployment. confidence: medium relevance: low - claim_id: domain_mismatched_asr_produces_misleading_wer role: complicates claim: Automatic ASR module benchmarking on conversational telephone speech corpora can produce highly misleading word error rates when the candidate ASR model's training domain does not match the conversational, disfluent register of the test data. source: §7, Table 2 evidence: An ESPnet LibriSpeech-trained ASR model evaluated on Switchboard Eval 2000 yields a WER of 342.2% and CER of 298%, far worse than domain-matched models like OWSM 3.1 (14.4% WER), illustrating a severe train/test domain mismatch effect rather than a genuine architecture comparison. confidence: high relevance: low - claim_id: transcript_quality_affects_downstream_tts_intelligibility role: supports claim: The choice of input transcript quality substantially affects downstream TTS intelligibility metrics independent of the TTS model itself, because more grammatically well-formed text is easier for TTS systems to synthesize accurately. source: §A.4, Table 8 evidence: Replacing ground-truth (disfluent, spontaneous) Switchboard transcripts with LLM-generated text responses as TTS input improves intelligibility across all tested TTS models, e.g. LJSpeech VITS WER drops from 16.1% to 6.0% under OWSM 3.1 hypotheses. confidence: high relevance: low - claim_id: dialogue_lm_sensitivity_to_asr_noise_varies_by_size role: complicates claim: Text dialogue response generators integrated into a cascaded spoken dialogue pipeline are not strongly degraded by upstream automatic speech recognition errors relative to using ground-truth transcripts, though smaller language models may be more sensitive to this noise than larger ones. source: §A.3, Table 7 evidence: When substituting OWSM 3.1 ASR transcripts for ground-truth transcripts as dialogue context, LLaMA-3.2-1B perplexity changes only slightly (48.2 to 47.5) while SmolLM v2-1.7B perplexity degrades substantially (113.7 to 285.1). confidence: medium relevance: low limitations: - The headline comparative analysis is conducted on a single dataset (11 hours of Switchboard Eval 2000) with only one example E2E system (Mini-Omni) and a 4-participant human pilot drawn from the author group; findings should be read as a toolkit demonstration rather than a generalizable benchmark. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.naacl-long.464 published_date: "2025-04-29" entry_date: '2026-07-27' year: 2025 venue: NAACL task: - TTS - evaluation architecture: [] relevance: medium evidence_role: - core_evidence - infrastructure current_role: active_evidence method_family: [] claims: - claim_id: multi_asr_voting_produces_aligned_data role: supports claim: Open-pipeline corpus construction using multi-ASR majority voting can produce high-quality aligned TTS data even when no single ASR model is reliable. source: §5.2, §C, Table 7 evidence: The ManaTTS pipeline achieved 97.98% chunk acceptance and a manual CER of 0.01 on 100 sampled chunks, despite relying on Persian ASRs with 10-30% CER on evaluation sets. confidence: high relevance: low - claim_id: open_corpus_absence_is_primary_bottleneck role: supports claim: The absence of open-access TTS corpora is a primary bottleneck for speech synthesis in low-resource languages. source: §2, Table 1, §6 evidence: All prior Persian TTS datasets are either not publicly available, synthetically generated, domain-limited (Quran-only), or under restrictive licenses; ManaTTS is the first to address all five criteria simultaneously. confidence: medium relevance: low - claim_id: high_mos_may_reflect_weak_baselines role: complicates claim: High MOS scores on a TTS system trained in a low-resource setting may reflect weak baselines rather than strong absolute quality. source: §5.1, Table 2 evidence: The proposed Tacotron2 model scores MOS 3.76, outperforming open-access Persian VITS and Glow-TTS baselines that score 1.68 and 1.34 respectively; the gap to natural speech (4.01) and to GT-spectrogram synthesis (3.86) is non-trivial. confidence: medium relevance: high - claim_id: forced_alignment_needs_mismatch_robustness role: refines claim: Forced alignment pipelines for TTS corpus construction require robustness to text-audio mismatch when sourcing from naturalistic recordings. source: §3.2.3, Figure 11 evidence: Standard alignment tools (e.g., Aeneas) assume exact text-audio correspondence; the custom gapped-interval alignment algorithm accepted 99.39% of chunks via Interval Search and handled mismatch cases via Gapped Search, critical for magazine-sourced data where speakers censor or self-correct. confidence: medium relevance: low limitations: - The MOS evaluation uses only five utterances and a single speaker, and comparisons are restricted to open-access Persian baselines that score far below acceptable naturalness; no comparison to current SOTA multilingual TTS systems is reported. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.naacl-long.619 published_date: "2025-04-29" entry_date: '2026-07-27' year: 2025 venue: NAACL task: - TTS architecture: - diffusion - transformer-enc-dec relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: - transformer_encoder_decoder - diffusion_generation claims: - claim_id: latent_space_diffusion_reduces_steps_without_degradation role: supports claim: Applying diffusion in a compact latent space rather than over the full signal substantially reduces the number of inference steps required without degrading output quality. source: §6, Table 3 evidence: ProSE achieves state-of-the-art PESQ and STOI on VoiceBank-DEMAND and TIMIT+MUSAN using only T=2 diffusion steps, outperforming baselines that require 6-50 steps. confidence: high relevance: low - claim_id: prior_regression_separation_reduces_artifacts role: supports claim: Separating prior generation by a diffusion model from regression-based reconstruction reduces artifacts caused by direct multi-step denoising in speech enhancement. source: §6, Figure 3 evidence: Waveform and mel-spectrogram comparisons show that DOSE introduces extraneous misaligned details absent in clean speech, whereas ProSE does not, because the regression model performs the actual mapping. confidence: medium relevance: low - claim_id: diffusion_se_generalization_not_uniform role: complicates claim: Diffusion-based speech enhancement methods do not uniformly generalize to real-world noise conditions unseen during training. source: §5, Table 4 evidence: Several baselines (DiffuSE, CDiffuSE, SGMSE, DR-DiffuSE, DOSE) show markedly lower subjective MOS in mismatched CHiME-4 conditions, suggesting condition-collapse or domain shift; ProSE is more robust but still incurs a gap from matched to mismatched conditions. confidence: medium relevance: high - claim_id: latent_diffusion_efficiency_holds_for_enhancement role: refines claim: The computational advantage of latent diffusion over waveform-space diffusion holds for speech enhancement tasks, not only for speech synthesis. source: §6, Table 3 evidence: ProSE uses fewer FLOPs (150.78G) than most diffusion-based SE baselines while achieving higher PESQ, attributable to running diffusion on a compact N=16, C'=256 latent rather than over the full mel-spectrogram or waveform. confidence: medium relevance: high limitations: - Comparisons are restricted to diffusion-based speech-enhancement baselines; no deterministic discriminative methods (e.g., Demucs, MANNER) are included, so the "state-of-the-art" claim applies specifically within the diffusion SE category. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.naacl-short.69 published_date: "2025-04-29" entry_date: '2026-07-27' year: 2025 venue: NAACL task: - TTS architecture: - flow-matching relevance: high evidence_role: - core_evidence - infrastructure current_role: active_evidence method_family: - flow_matching_generation claims: - claim_id: multilingual_training_improves_related_low_resource_langs role: supports claim: Multilingual training on typologically related low-resource languages improves TTS objective quality over monolingual training for all participating languages. source: §5, Table 2 evidence: Multilingual models consistently outperform monolingual baselines across F0-RMSE, MCD, PESQ, and VUV F1 for all four speakers (Ojibwe JJ, NJ; Mi'kmaq MJ; Maliseet AT); MOS for Ojibwe JJ improves from 2.71 (monolingual) to 3.42 (multilingual FNet). confidence: high relevance: high - claim_id: attention_free_mixers_match_self_attention_half_memory role: supports claim: Attention-free sequence mixing architectures can match self-attention performance in flow-matching TTS while halving GPU memory requirements. source: §5, Tables 2-3 evidence: Hydra achieves objective metrics close to self-attention MatchaTTS across all four speakers, and attention-free models consume 2.3-2.5G vs. 4.6G GPU memory at batch size 400 (float32). confidence: medium relevance: low - claim_id: subquadratic_complexity_no_latency_gain_at_small_scale role: complicates claim: Subquadratic complexity in attention-free architectures does not translate to inference latency gains in small TTS models. source: §5, Table 3 evidence: Self-attention has higher throughput than Mamba2, Hydra, and FNet in float32 on an A100 at batch size 400, because the model is small enough that architectural complexity advantages are not observable. confidence: medium relevance: low - claim_id: mos_unreliable_for_indigenous_communities role: complicates claim: Standard MOS-based subjective evaluation protocols can produce unreliable results when applied to Indigenous language TTS due to cultural norms around negative judgement. source: §5 evidence: One of two Ojibwe raters assigned the highest score (5/5) to all voices including synthetic ones; MOS surveys could not be completed for Mi'kmaq and Maliseet due to lack of available raters; the authors argue the task paradigm itself is culturally misaligned. confidence: medium relevance: high limitations: - The custom datasets are not publicly released due to lack of consent from voice donors, preventing replication; MOS evaluation is available only for one Ojibwe speaker (two raters), making subjective assessment effectively impossible for three of the four speakers. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.iwsds-1.11 published_date: "2025-05-01" entry_date: '2026-07-27' year: 2025 venue: IWSDS task: - SCA architecture: [] relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: [] claims: - claim_id: ssl_representations_capture_paralinguistic_attitude role: supports claim: Self-supervised speech representations capture sufficient prosodic information to classify paralinguistic attitudes from acoustic input alone, without linguistic features. source: §4.1, Table 4, Table 5 evidence: A frozen HuBERT-large model using layer 12 embeddings averaged over time achieves macro-F1 of 0.909 on a four-class attitude task in Japanese reading speech, exceeding human listener performance of 0.829 on the same data. confidence: high relevance: high - claim_id: reverberation_more_persistent_than_noise role: complicates claim: Environmental reverberation poses a more persistent challenge than additive noise for paralinguistic recognition models, and is not adequately addressed by speech enhancement postprocessing. source: §4.3, Table 7 evidence: Under noisy conditions, MP-SENet speech enhancement recovers macro-F1 from 0.625 to 0.844; under reverberant conditions, the same model raises F1 only from 0.449 to 0.492, attributed to the difficulty of recovering prosodic structure from reverberant speech. confidence: medium relevance: low - claim_id: augmentation_does_not_close_inference_time_gap role: complicates claim: Data augmentation with noise and reverberation during training does not close the performance gap when those conditions appear at inference time. source: §4, §4.3, Table 7 evidence: Despite four-fold augmentation using noise (DEMAND, MUSAN, FSD50K) and room impulse responses (BIRD) at varying SNR levels, macro-F1 drops from 0.912 (clean) to 0.625 (noisy) and 0.449 (noisy-reverberant) on held-out test conditions. confidence: medium relevance: low - claim_id: reading_speech_may_overstate_real_world_performance role: complicates claim: Paralinguistic attitude recognition systems trained on reading-speech datasets may overstate real-world spoken dialogue performance, where speech production is more spontaneous and less controlled. source: §3, §5 evidence: The dataset consists entirely of scripts read with intended attitudes by crowd workers and actors; the authors identify generalisation to naturally-occurring speech directed at dialogue systems as unresolved future work. confidence: low relevance: high limitations: - The corpus is a Japanese reading-speech dataset where speakers deliberately produce each attitude; it does not reflect the ambiguity or variability of spontaneous human-machine dialogue speech, and reverberation robustness remains largely unaddressed by tested enhancement postprocessing. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.iwsds-1.27 published_date: "2025-05-01" entry_date: '2026-07-27' year: 2025 venue: IWSDS task: - SCA architecture: [] relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: [] claims: - claim_id: no_shared_benchmarks_impedes_progress_measurement role: complicates claim: Systematic progress measurement in turn-taking modeling is impeded by the absence of shared evaluation benchmarks. source: §7.1 evidence: 72% of surveyed papers do not compare their methods with prior work; one third of EOU detection papers use no public corpus, making cross-study comparison unreliable. confidence: high relevance: low - claim_id: multimodal_cues_improve_turn_taking_prediction role: supports claim: Combining multimodal cues (prosodic, lexical, and visual) consistently improves turn-taking prediction over single-modality approaches. source: §4.1 evidence: Multiple independent surveyed works demonstrate that joint use of prosodic and word features outperforms either in isolation, consistent with prior findings on the additive effect of turn-taking cues across human communication. confidence: medium relevance: high - claim_id: vap_models_unify_eou_and_backchannel_prediction role: supports claim: Voice Activity Projection models provide a self-supervised framework that unifies end-of-turn and backchannel prediction in continuous spoken dialogue. source: §4.1, §5 evidence: VAP models evaluated zero-shot across shift prediction, hold prediction, upcoming backchannel, and backchannel-vs-turn-shift classification outperform earlier silence-threshold and IPU-based baselines in multiple reviewed works. confidence: medium relevance: low - claim_id: llms_insufficient_standalone_turn_taking_detectors role: complicates claim: Large language models are insufficient as standalone turn-taking detectors in spoken dialogue. source: §4.1 evidence: A surveyed study demonstrates that LLMs are ineffective at detecting mid-utterance turn initiation points without integration of acoustic signals. confidence: medium relevance: low - claim_id: cross_lingual_transfer_requires_segmentation_alignment role: complicates claim: Cross-lingual transfer for turn-taking models requires alignment of speech segmentation criteria, not merely multilingual training data. source: §4.1, §7.3 evidence: VAP models show poor zero-shot cross-lingual transfer; alignment of IPU silence threshold definitions across datasets is identified as a prerequisite for valid multilingual evaluation. confidence: medium relevance: low limitations: - The literature search focuses on post-2021 work from major conference proceedings and may underrepresent preprints and workshop papers outside the primary venues searched; the multi-party conversation section is acknowledged as sparse. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.americasnlp-1.1 published_date: "2025-05-04" entry_date: '2026-07-27' year: 2025 venue: AmericasNLP 2025 task: - TTS architecture: - transformer-enc-dec - GAN relevance: high evidence_role: - core_evidence - infrastructure current_role: active_evidence method_family: - transformer_encoder_decoder - gan_adversarial_generation claims: - claim_id: transfer_learning_enables_low_resource_indigenous_tts role: supports claim: Transfer learning from a closely related language can enable intelligible TTS for an indigenous language with fewer than 5 hours of training data. source: §4.1, §5.1 evidence: Tacotron 2 fine-tuned from a Latin American Spanish checkpoint on 4h37m of Shipibo-Konibo speech achieves Intelligibility Rate 88.56% and MOS 4.01, both exceeding the natural speech reference recordings. confidence: medium relevance: high - claim_id: subjective_evaluation_necessary_without_reference_audio role: supports claim: Subjective evaluation by native speakers is necessary when objective metrics such as PESQ and POLQA cannot be applied due to absent high-quality reference recordings. source: §5.3 evidence: No high-quality studio reference existed for Shipibo-Konibo, so the authors relied entirely on MOS and Intelligibility Rate ratings from 26 native-speaker evaluators rather than reference-based objective metrics. confidence: high relevance: high - claim_id: speaker_pronunciation_drift_confounds_natural_vs_synthetic_comparison role: complicates claim: Pronunciation drift in a native speaker living outside the language community can cause natural speech to score lower on intelligibility than synthetic speech on standardised evaluation. source: §5.4 evidence: The single native recording speaker's years of residence in Lima introduced pronunciation variations into the reference recordings, which the authors offer as the explanation for synthetic speech (MOS 4.01, IR 88.56%) outscoring natural speech (MOS 3.75, IR 83.45%). confidence: medium relevance: high - claim_id: gan_vocoder_converges_fast_from_similar_pretraining_distribution role: supports claim: GAN-based vocoders initialized from a universal pretrained model converge rapidly when the target speaker is similar to the pretraining distribution. source: §4.2, §5.2 evidence: HiFi-GAN initialized from a pretrained universal female voice model converged after only 34 epochs when fine-tuned on the single-speaker Shipibo-Konibo corpus. confidence: medium relevance: low limitations: - Evaluation relied entirely on subjective metrics (no PESQ/POLQA due to lack of reference data); subjectivity and evaluator bias are acknowledged. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2507.06235' published_date: "2025-05-20" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: [] relevance: high evidence_role: - core_evidence current_role: active_evidence method_family: [] claims: - claim_id: f0_f1_predict_perceived_cuteness role: supports claim: Fundamental frequency and lower formant frequencies are primary acoustic predictors of perceived cuteness in synthetic voices. source: §4.6.1 evidence: Phase 1 Spearman correlations show strong positive relationships between kawaii ratings and F0 (rs=0.89) and F1 (rs=0.74), while F2 and F3 are non-significant. confidence: high relevance: low - claim_id: tts_cuteness_manipulation_fails_on_professional_voices role: complicates claim: Acoustic frequency manipulations that amplify cuteness perceptions in generative TTS voices do not transfer to naturally recorded or professionally processed voices. source: §5.4, Table 6 evidence: The three-semitone F0/formant shift that improved kawaii ratings for TTS voices significantly reduced kawaiiness for professionally recorded game character voices (Mann-Whitney U test, p<0.001), attributed to processing artefacts and possible ceiling effects. confidence: high relevance: low - claim_id: automated_vocoder_shifting_matches_manual_editing_on_primary_but_not_secondary_attributes role: complicates claim: Automated vocoder-based pitch and formant shifting matches manual audio editing for cuteness perception but diverges on secondary perceptual attributes. source: §4.6.2, Table 3, Table 4 evidence: Wilcoxon signed-rank tests found no significant kawaii-rating differences across Cubase, Legacy-STRAIGHT, and WORLD manipulation methods, but significant differences emerged on humanlikeness, animal-likeness, trustworthiness, and excitedness ratings for several voices. confidence: medium relevance: high - claim_id: social_desirability_outpredicts_frequency_for_cuteness role: supports claim: Perceived cuteness of synthetic voices is more strongly predicted by social desirability attributes than by acoustic frequency measures in isolation. source: §5.4, Table 7 evidence: GEE analysis across game character voices found favorability, humanlikeness, familiarity, and trustworthiness as significant predictors of kawaiiness, with effect sizes exceeding those of frequency correlations alone. confidence: medium relevance: high limitations: - All four study phases recruited exclusively Japanese adults, so generalisability of the kawaii vocalics manipulation to non-Japanese listeners or other languages is untested. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2505.15772' published_date: "2025-05-21" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS - evaluation architecture: [] relevance: high evidence_role: - core_evidence current_role: active_evidence method_family: [] claims: - claim_id: automated_multimodal_annotation_exceeds_human_interrater_consistency role: supports claim: Automated multimodal annotation can match human accuracy while substantially outperforming human annotation in inter-rater consistency for emotional speech labeling. source: §2.3, Table 1 evidence: MIKU-PAL achieves Fleiss kappa of 0.93-0.95 on IEMOCAP and MELD across five independent annotation runs, compared to 0.40-0.43 for human annotators, at roughly $0.50/hour versus $10/hour for manual labeling. confidence: high relevance: high - claim_id: fine_grained_emotion_taxonomy_improves_downstream_tts_expressiveness role: supports claim: Expanding emotional speech training data beyond conventional 6-category taxonomies to psychologically grounded fine-grained categories improves downstream TTS emotion expressiveness. source: §3.2, Table 3 evidence: Fish-Speech fine-tuned on MIKU-EmoBench (26 categories) achieves MOS 4.12 and emotion similarity 0.92, outperforming fine-tuning on IEMOCAP and MELD (MOS 4.01/4.00, emotion similarity 0.89 each). confidence: medium relevance: high - claim_id: standard_emotion_categories_inflate_apparent_error_rates role: complicates claim: Standard emotional speech benchmark categories introduce annotation ambiguity that inflates apparent system error rates in automatic labeling. source: §2.3, Figure 4 evidence: MIKU-PAL accuracy on IEMOCAP rises from approximately 58.6% to approximately 75% when 'frustration' and 'neutral' categories, which lack a distinct basis in established psychological classification, are excluded from scoring. confidence: medium relevance: high - claim_id: visual_modality_improves_speech_emotion_classification role: supports claim: Visual modality information contributes substantially to speech emotion classification accuracy in multimodal annotation systems. source: §2.1 evidence: Ablation on MELD shows a 25.6% accuracy improvement when vision preprocessing (face detection and active speaker identification) is included relative to audio-only analysis. confidence: medium relevance: high limitations: - MIKU-PAL's annotation quality depends on Gemini 2.0 Flash, a proprietary model whose availability, versioning, and cost structure may change, limiting reproducibility over time. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2506.23049' published_date: "2025-06-29" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - SCA architecture: - autoregressive-LM relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - autoregressive_generation claims: - claim_id: cascaded_pipelines_outperform_e2e_on_practical_benchmarks role: supports claim: Cascaded ASR+LLM+TTS pipelines outperform end-to-end speech models on practical spoken task benchmarks when a high-quality instruction-tuned text LLM serves as the reasoning backbone. source: §III.A, Table I evidence: AURA with Whisper+LLaMA-3.3-70B achieves 92.75% on VoiceBench OpenBookQA versus 25.93% (Moshi) and 26.59% (Mini-Omni2) for end-to-end speech models, and 83.52% for Kimi-Audio. confidence: high relevance: high - claim_id: retrieval_augmentation_improves_spoken_qa role: supports claim: Retrieval augmentation via web search at inference time improves spoken question answering on knowledge-intensive multiple-choice tasks without retraining. source: §III.A evidence: Prompting AURA (OWSM ASR) to perform at least one web search before answering raises OpenBookQA accuracy from 87.69% to 89.23%. confidence: medium relevance: low - claim_id: react_reasoning_enables_reliable_multiturn_tool_use role: supports claim: ReAct-style interleaved reasoning and action enables reliable tool selection and goal completion in multi-turn spoken conversational agents. source: §III.B, Table II evidence: AURA achieves 90% overall task success and satisfaction scores above 4.0 across 30 human-evaluated tasks spanning compound goals requiring calendar, email, contact, and web search APIs. confidence: medium relevance: high - claim_id: dst_remains_unreliable_even_with_large_llms role: complicates claim: Dialog state tracking in spoken task-oriented systems remains far from reliable even with large instruction-tuned LLMs applied without fine-tuning. source: §III.C, Table III evidence: AURA achieves 28.76% JGA on SpokenWOZ, improving over the best prior baseline by 3.1 points but remaining well below full accuracy. confidence: medium relevance: low limitations: - The human evaluation of multi-turn task success uses a single evaluator (a co-author), which introduces potential bias and limits the reliability of reported success rates. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2507.00808' published_date: "2025-07-02" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - transformer-enc-dec relevance: high evidence_role: - core_evidence current_role: active_evidence method_family: - transformer_encoder_decoder claims: - claim_id: iterative_natural_language_feedback_can_progressively_refine_the role: supports claim: Iterative natural language feedback can progressively refine the speaking style of synthesized speech without accumulating naturalness degradation. source: §4.2, §4.4, Figure 4, Figure 6 evidence: Over three interaction sessions, Iterative (ours) significantly outperformed the Identical baseline on a 5-point style refinement MOS, and naturalness MOS showed no significant difference between Iterative and the oracle condition across all style groups. confidence: high relevance: high - claim_id: global_speech_embedding_based_conditioning_cannot_accurately_reflect role: complicates claim: Global speech embedding-based conditioning cannot accurately reflect fine-grained positional or linguistic instructions in expressive TTS. source: §5.1, §5.2, Table 4 evidence: Low-scoring examples in the test set contained directions targeting specific word positions ("at the beginning", "at the end", "for the part of...") or linguistic modifications ("hold your breath", "place just a slight pause between words"), which the speech embedding manipulation approach could not handle. confidence: high relevance: low - claim_id: semantic_similarity_of_style_direction_text_not_exact role: supports claim: Semantic similarity of style direction text, not exact wording, governs how well listeners perceive style refinement as aligned with the instruction. source: §4.3, Figure 5 evidence: In the style refinement accuracy evaluation, directions semantically similar to the one used for refinement (Random Similar) yielded scores comparable to the Matched condition, while semantically dissimilar directions scored significantly lower across all style groups. confidence: high relevance: high - claim_id: holistic_subjective_evaluation_scales_may_not_adequately_capture role: complicates claim: Holistic subjective evaluation scales may not adequately capture fine-grained stylistic alignment in iterative TTS refinement tasks. source: §4.2, §5.3 evidence: Even the Actor-Guided oracle condition scored around 3 out of 5 on the iterative style refinement MOS, which the authors attribute to the evaluation task not fully discriminating subtle style differences; similar evaluation difficulties have been noted in text-to-image/video generation research. confidence: high relevance: high limitations: - All training and evaluation data is proprietary in-house Japanese speech from two voice actors. No public dataset is used, and no results are reported outside this setup. Reproducibility and generalization are untested. - 'The style refiner is speaker-dependent; the authors plan to extend to speaker-independent operation as future work. The current model refines only paralinguistic information (speaking style via global embeddings), not linguistic content, so instructions involving pauses, stress, or pitch accent placement cannot be followed. The directions cover only two of four practical categories from actual recording sessions (paralinguistic and text-expressible linguistic instructions), omitting demonstrative and gestural instructions entirely. The evaluation task design is also noted as an open problem: the relatively low absolute scores even under oracle conditions suggest that existing MOS paradigms do not cleanly measure this type of fine-grained iterative alignment, and more sensitive evaluation methods are needed.' caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2507.01611' published_date: "2025-07-02" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - GAN relevance: high evidence_role: - core_evidence current_role: active_evidence method_family: - gan_adversarial_generation claims: - claim_id: explicit_envelope_modeling_enables_pitch_extrapolation role: supports claim: Explicit spectral envelope modeling in a neural vocoder enables reliable pitch modification beyond the training frequency range, where purely neural waveform predictors fail. source: §V.B, §V.D, Tables II, VI, Fig. 6 evidence: QHARMA-GAN maintains f0 RMSE of 0.08-0.11 Hz across pitch-scale factors from 0.5x to 2x on VCTK, while Vocos fails catastrophically at 0.65-0.67 Hz. confidence: high relevance: low - claim_id: signal_processing_priors_reduce_vocoder_data_requirements role: supports claim: Incorporating classical signal-processing constraints into a neural vocoder reduces data requirements without sacrificing subjective quality. source: §V.D, Table VII evidence: QHARMA-GAN trained on 919 LJSpeech utterances achieves MOS 3.85 compared to HiFi-GAN's 3.53 under the same data constraint, with HiFi-GAN overfitting on insufficient data. confidence: medium relevance: high - claim_id: pesq_diverges_from_mos_for_hybrid_vocoders role: complicates claim: Hybrid vocoders that analytically reconstruct phase can achieve higher subjective naturalness than end-to-end neural vocoders despite lower PESQ scores, suggesting objective spectral distance metrics do not fully capture perceptual frequency smoothness. source: §V.B, Table III evidence: QHARMA-GAN scores MOS 4.21 on VCTK versus HiFi-GAN's 4.08, while PESQ is comparable (3.14 vs 3.14) and both trail QHM's PESQ of 3.45 despite QHM scoring the lowest subjective MOS. confidence: medium relevance: high - claim_id: vuv_detection_bottleneck_for_pitch_modification role: complicates claim: V/UV detection errors are a critical bottleneck for hybrid pitch-modification vocoders, particularly at extreme pitch-raising factors. source: §V.B, Table IV evidence: QHARMA-GAN outperforms WORLD in pitch lowering (MOS 3.01 vs 2.98 at rho=0.5x) but underperforms WORLD in pitch raising (2.72 vs 2.82 at rho=2x), attributed to V/UV misclassification during upward pitch shifts. confidence: medium relevance: high limitations: - Evaluations use proprietary metrics (UTMOS) alongside standard ones; no listening test confidence intervals are reported for the modification experiments, making statistical significance of cross-method differences difficult to assess. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2507.02176' published_date: "2025-07-02" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - evaluation architecture: [] relevance: high evidence_role: - core_evidence current_role: active_evidence method_family: [] claims: - claim_id: asv_embeddings_miss_dynamic_identity_markers role: supports claim: ASV embeddings encode static anatomical speech features but systematically fail to represent dynamic behavioral identity markers such as rhythm and timing patterns. source: §3.2, Figure 1 evidence: Lasso regression predicting handcrafted features from seven ASV models' embeddings shows high r² for mean pitch, HNR, shimmer, and α-ratio, but near-zero r² for speech rate, voiced/unvoiced segment lengths, and pitch standard deviation. confidence: high relevance: low - claim_id: eer_confounded_by_non_identity_factors role: complicates claim: EER-based speaker similarity measurements in speech synthesis evaluation are susceptible to confounding factors unrelated to voice identity, which can invalidate comparisons between synthesis systems. source: §3.3, Table 2 evidence: Duration-sorting same-speaker utterances depresses EER from an expected 50% to 30-39% across all ASV models tested; SNR 20dB noise reduces EER to 15-38%; equalization mismatches cause near-total failure in GE2E, all correctable via re-equalization and duration matching. confidence: high relevance: high - claim_id: rhythm_identity_requires_phoneme_duration_distributions role: supports claim: Characterizing speaker rhythm for identity assessment requires modeling phoneme-duration distributions rather than aggregate measures such as mean speech rate. source: §3.4, Figure 2, Table 3 evidence: U3D Wasserstein distances clearly separate same-speaker pairs (avg. 2.15) from nearest-by-speech-rate pairs (18.40) and random pairs (21.53), while many L2-ARCTIC speakers sharing similar syllable rates show markedly different voiced-segment duration patterns. confidence: high relevance: low - claim_id: ssl_units_substitute_for_phoneme_labels_in_rhythm_analysis role: refines claim: Self-supervised speech unit representations serve as language-agnostic substitutes for phoneme labels in rhythm analysis, avoiding the need for forced alignment. source: §3.4, Table 3 evidence: U3D using HuBERT-derived unsupervised clusters achieves Wasserstein distance separation between speaker conditions (2.15 same, 18.40 nearest, 21.53 random) nearly identical to forced-alignment-based phoneme rhythm analysis (2.48, 18.37, 24.43). confidence: medium relevance: low limitations: - U3D is validated as a discriminative metric (separating speaker pairs) but has not been validated against human perceptual judgments of rhythm similarity. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2507.04817' published_date: "2025-07-07" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - VC architecture: - GAN relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: - gan_adversarial_generation claims: - claim_id: explicit_prosody_conditioning_matches_larger_vc_systems role: supports claim: Explicit, interpretable prosodic conditioning in voice conversion can match the intelligibility and speaker similarity of systems using implicit or disentangled representations, at a fraction of the parameter cost. source: §5.1.1, Table 2 evidence: Fast-VGAN (~3.2M params) achieves WER of 0.000% and cosine similarity 0.648, competitive with ControlVC (~20M params, WER 0.089%, similarity 0.652) and superior to HiFi-VC (~14M params, WER 2.857%). confidence: high relevance: high - claim_id: joint_prosodic_adaptation_degrades_naturalness_and_similarity role: complicates claim: Adapting target-speaker-specific prosodic parameters (pitch ambitus, speech rate) during voice conversion does not reliably improve perceived speaker identity, and combining multiple prosodic adaptations simultaneously degrades both naturalness and similarity. source: §5.2.1, Table 4 evidence: When ambitus and speech rate dilation are applied jointly, subjective MOS naturalness drops from 3.63 to 3.13 and speaker similarity from 3.47 to 2.99 compared to baseline conversion without prosodic parameter adaptation. confidence: medium relevance: high - claim_id: neutral_to_expressive_transfer_without_expressive_training_data role: supports claim: Neutral-to-expressive speech resynthesis can be achieved by applying prosodic contours from expressive reference recordings to a model trained exclusively on neutral speech, without requiring expressive training data. source: §4.3, §5.1.3, Table 3 evidence: Fast-VGAN trained only on neutral Expresso speaker utterances achieves 0.867 cosine speaker similarity on expressive resynthesis across confused, happy, and sad conditions, with WER comparable to real recordings. confidence: medium relevance: high - claim_id: extreme_prosodic_transforms_degrade_bell_shaped role: complicates claim: Extreme prosodic transformations in GAN-based voice conversion degrade naturalness and speaker similarity, with performance falling off on both sides of the neutral setting in a bell-shaped pattern. source: §5.2.2, Figure 3 evidence: MOS scores for naturalness and speaker similarity peak at unmodified settings and decline for both compression and expansion of duration, ambitus, and F0 transposition. confidence: medium relevance: high limitations: - Speaker generalisation is restricted to seen speakers, as speaker identity is encoded as a learned lookup embedding rather than a d-vector or x-vector, so the model cannot convert to unseen target speakers without retraining. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2507.06116' published_date: "2025-07-08" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: auxiliary_model_classification_improves_system_level_mos_accuracy role: supports claim: Multi-task learning with an auxiliary synthesis-model classification objective can improve absolute MOS prediction accuracy at the system level by encouraging the model to learn discriminative acoustic quality features. source: §III.C, Table 2 evidence: The MoE system with auxiliary classification achieves system-level MSE 0.056 (rank 1 of seven teams), a 21% reduction in MSE over the strongest single-objective baseline (MSE 0.071). confidence: high relevance: high - claim_id: system_level_gains_dont_transfer_to_utterance_level role: complicates claim: Improvements in system-level MOS prediction accuracy do not reliably transfer to utterance-level prediction performance. source: §IV, Table 1 evidence: Despite ranking first on system-level MSE, the proposed model places 3rd-6th across utterance-level metrics; the authors identify disjoint train/test annotator sets as a primary cause. confidence: high relevance: high - claim_id: system_utterance_gap_reflects_different_task_requirements role: refines claim: The performance gap between system-level and utterance-level MOS prediction reflects fundamentally different task requirements, not a data-scaling problem alone. source: §IV evidence: Doubling training data from 400 to 800 samples and introducing the MoE architecture yields only marginal utterance-level gains, while system-level gains are large, which the paper attributes to micro-feature variability and rater-specific bias rather than data volume. confidence: medium relevance: high - claim_id: high_system_level_correlation_doesnt_imply_utterance_level_correlation role: complicates claim: High correlation between predicted and actual MOS at the system level does not imply high correlation at the utterance level for the same model. source: §IV, Tables 1-2 evidence: The proposed model achieves LCC 0.978 on system-level evaluation but only LCC 0.811 on utterance-level evaluation, with similarly large gaps in SRCC and KTAU. confidence: high relevance: high limitations: - The training and test annotator pools are completely disjoint (raters 0-9 for training, raters 10-19 for testing), meaning utterance-level results measure generalisation to unseen rater subjectivity rather than a standard train/test split. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2506.09874' published_date: "2025-07-10" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - flow-matching relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - flow_matching_generation claims: - claim_id: joint_generation_beats_post_hoc_mixing role: supports claim: Joint generative modelling of speech and environmental audio produces more naturally integrated output than post-hoc mixing of separately generated speech with background sounds. source: §4.2 evidence: A/B tests with 20+ raters per sample showed 87.89% preference for UmbraTTS over F5-TTS plus post-hoc background mixing across 30 AudioSet samples. confidence: high relevance: high - claim_id: self_supervised_extraction_sufficient_for_environmental_tts role: supports claim: Self-supervised extraction of speech and background audio from unlabeled mixed recordings is sufficient to train environmentally conditioned TTS to state-of-the-art performance. source: §3, §4.1, Table 1 evidence: UmbraTTS uses VAD-based silence extraction and learned source separation applied to AudioSet to construct training triplets without manual annotation, achieving best WER and FAD among all baselines on AudioCaps. confidence: medium relevance: high - claim_id: flow_matching_beats_diffusion_for_environmental_tts role: supports claim: Flow matching outperforms diffusion-based architectures in environmentally aware speech synthesis on both automatic and human evaluation metrics. source: §4.1, §4.2, Tables 1-3 evidence: UmbraTTS (CFM-based) achieves WER 6.89% and FAD 4.14 versus VoiceDiT (diffusion, 7.09%, 4.60) and VoiceLDM (diffusion, 10.39%, 5.56) on AudioCaps, with 78.54-81.91% human preference over VoiceLDM. confidence: medium relevance: high - claim_id: explicit_ser_control_needed_for_background_balance role: complicates claim: Controlling speech-to-environment balance in jointly generated audio scenes requires an explicit scalar conditioning parameter; standard text and speaker conditioning are insufficient for this degree of output control. source: §3, §4.2 evidence: UmbraTTS introduces a dedicated SER signal injected via adaLN-zero; a separate controlled listening test confirms 96.6% of participants correctly distinguish the intended background volume direction in pairwise comparisons. confidence: medium relevance: high limitations: - The evaluation protocol is borrowed from VoiceLDM and covers only the AudioCaps and MusicCaps test subsets, so generalisation to other environmental conditions is untested. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2507.09310' published_date: "2025-07-12" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - VC - TTS architecture: - VAE relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - variational_latent_generation claims: - claim_id: implicit_style_modeling_preserves_speech_style role: supports claim: Implicit style modeling via an auxiliary reconstruction loss can preserve speaking style characteristics in voice conversion without requiring explicit acoustic feature extraction. source: §3.3, §3.4, Table 2 evidence: A style reconstruction loss using a frozen Lombard/neutral classifier achieves SIIB scores comparable to explicit conditioning on f0, mgc0, and mgc1 across male and female target speakers, with no statistically significant differences in most conditions. confidence: medium relevance: high - claim_id: explicit_feature_effectiveness_gender_dependent role: complicates claim: The effectiveness of explicit prosodic features for style-preserving voice conversion depends on speaker gender, limiting the generalizability of feature-based conditioning approaches. source: §3.4, Table 2 evidence: F0 conditioning improves SIIB for male target speakers but is detrimental for female target speakers at both SNR-1 and SNR-3; spectral tilt and energy (mgc0+mgc1) are beneficial across genders. confidence: medium relevance: low - claim_id: speaker_similarity_perceptible_in_noise role: supports claim: Speaker identity remains perceptible to listeners in noisy conditions, confirming that speaker similarity is a relevant evaluation dimension for intelligibility-enhancing voice conversion. source: §2.3, Figure 1, §3.4, Figure 6 evidence: Pilot study results show that listeners maintain consistent similarity rankings between speakers at SNR-1 and SNR-3 noise levels, and the implicit conditioning VC model preserves speaker similarity comparably to the CopyCat baseline. confidence: medium relevance: high - claim_id: fusion_inconsistent_improvement role: complicates claim: Combining implicit and explicit style conditioning does not consistently improve intelligibility over either approach alone, and can degrade performance in some configurations. source: §3.4, Figure 5, Table 2 evidence: Fusion of Ls with individual explicit features (e.g., Ls+f0 for female targets) reduces SIIB compared to Ls alone; only Ls+mgc0+mgc1 reliably improves over Ls in most conditions, and gains are inconsistent across genders. confidence: medium relevance: high limitations: - The evaluation is restricted to a single small dataset (54 speakers, 100 short utterances each) with only two VC target speakers, limiting how far the intelligibility and speaker-similarity findings generalize. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2506.18296' published_date: "2025-07-15" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS - VC - evaluation architecture: [] relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: [] claims: - claim_id: nonanonymous_corpora_enable_rigorous_speaker_similarity_eval role: supports claim: Non-anonymous speaker corpora with publicly identifiable voices enable more rigorous subjective evaluations of speaker similarity in TTS and VC systems. source: §1, §3.1 evidence: JIS assigns stage names to 169 Japanese live idol speakers, allowing experiment designers to recruit listeners familiar with the speakers, enabling discrimination of subtle voice characteristics that anonymous corpus evaluations cannot capture. confidence: high relevance: high - claim_id: automatic_mos_underestimates_nonprofessional_quality role: complicates claim: Automatic MOS predictors trained on TTS-generated speech may underestimate audio quality when applied to recordings of non-professional speakers, even under studio conditions. source: §4.2.1 evidence: JIS Speech A (studio-recorded) achieves a mean UTMOS of 3.4 compared to JVS parallel100's 3.7, with the gap attributed to linguistic mismatch in UTMOS training data and speech hesitations inherent to non-professional speakers rather than recording quality differences. confidence: medium relevance: high - claim_id: speaking_style_shifts_speaker_embeddings role: supports claim: Speaking style and communicative context introduce systematic variation in speaker embeddings that interacts with speaker identity, presenting a challenge for style-robust speaker representation. source: §4.2.2 evidence: ECAPA-TDNN embeddings of JIS speakers show that specific speaking styles (energetic post-performance greetings, intimate photo-session speech) produce cross-speaker clusters in t-SNE, partially overriding individual speaker identity, while speech expressing personal individuality is more dispersed. confidence: medium relevance: low limitations: - The corpus is small (17 hours, 169 speakers) and its non-anonymous design is restricted to non-commercial basic research, which is a general corpus limitation rather than one specific to its evaluation-methodology contribution. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2507.14988' published_date: "2025-07-20" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - flow-matching relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: - flow_matching_generation claims: - claim_id: grpo_discrete_upstream_subcomponent_gain role: supports claim: Applying GRPO to a discrete upstream subcomponent of a distilled TTS pipeline achieves perceptual metric gains at substantially lower computational cost than whole-system RL. source: §3.2.2, §4.1, Table 1 evidence: Duration predictor GRPO fine-tuning uses 4-step student inference for reward computation and converges in 1.5K steps; on Seed-TTS-en, WER improves from 3.750 (no RL) to 1.752 and SIM from 0.672 to 0.698 without modifying the synthesis model. confidence: high relevance: high - claim_id: duration_prediction_bottleneck role: supports claim: Duration prediction quality independently constrains intelligibility and speaker similarity in non-autoregressive TTS regardless of synthesis model capacity. source: §A.1, Table 3 evidence: With the same flow-matching generator, WER on Seed-TTS-en ranges from 1.723 (best-of-8 oracle duration sampling) to 3.750 (unoptimized predictor) to 1.752 (RL-optimized), establishing duration as the primary bottleneck. confidence: high relevance: high - claim_id: distillation_reduces_prosodic_diversity role: complicates claim: Distribution matching distillation in flow-matching TTS reduces prosodic diversity disproportionately relative to acoustic quality, concentrating degradation in temporal and structural dimensions. source: §3.3.1, §4.3.3, Table 1, Figure 3 evidence: Step reduction from 32 to 4 lowers pitch CV-f0 by 30.3% (from 0.666 to 0.464) while mean-centered UTMOS distributions remain consistent across teacher and student, showing diversity loss is prosodic rather than spectral. confidence: high relevance: high - claim_id: best_of_n_practical_ceiling role: refines claim: Best-of-N sampling over discrete duration predictions establishes a practical performance ceiling for RL-based duration optimization in non-autoregressive TTS. source: §A.1, Table 3 evidence: Best-of-8 duration sampling achieves WER 1.723 and SIM 0.724 on Seed-TTS-en; the RL-optimized system reaches WER 1.752 and SIM 0.698 in a single pass, approaching but not fully closing the gap to this oracle upper bound. confidence: medium relevance: low - claim_id: teacher_guided_sampling_restores_diversity role: supports claim: A hybrid inference strategy using a high-diversity teacher for early denoising and a distilled student for acoustic refinement can restore prosodic diversity without sacrificing metric optimization gains. source: §3.3.2, §4.3.3, Table 1 evidence: Teacher-guided sampling (t_switch=0.25, 16 steps) recovers CV-f0 to 0.593 (89.1% of teacher diversity) while achieving WER 1.738 and SIM 0.699, outperforming the teacher on both objective metrics at 1.8x faster RTF. confidence: high relevance: low limitations: - Reward signals used for RL optimization are limited to WER and speaker similarity; explicit naturalness or prosody reward metrics are not explored, leaving the reward design incomplete relative to full perceptual quality. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2507.16875' published_date: "2025-07-22" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - flow-matching relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - flow_matching_generation claims: - claim_id: duration_strategy_language_dependent_tradeoffs role: complicates claim: Duration prediction strategy choice in zero-shot TTS presents language-dependent trade-offs between intelligibility and speaker similarity that are not consistent across multilingual settings. source: §5, Table 3, Table 5 evidence: The speaker-prompted (PFlow-style) predictor improves Tamil intelligibility (WER 0.295 vs. 0.312) and speaker similarity (SMOS 4.56 vs. 4.41) but degrades Marathi intelligibility substantially (WER 0.365 vs. 0.224 for infill), while Bengali human evaluation favours infill on both dimensions. confidence: medium relevance: high - claim_id: speaker_prompted_duration_no_forced_alignment role: supports claim: Speaker-conditioned audio prompting enables duration prediction without forced alignment, yielding competitive or superior speaker similarity in low-resource multilingual TTS. source: §3.4.2, §5.1, Table 4, Table 5 evidence: The 84M-parameter speaker-prompted predictor, conditioning on a 3-second mel spectrogram via cross-attention, achieves higher Sim-o scores than the infill predictor for Tamil, Telugu, and Bengali, and substantially higher human SMOS for Hindi (4.14 vs. 3.57), without using explicit duration context at inference. confidence: medium relevance: high - claim_id: pooled_corpora_quality_variance_requires_filtering role: complicates claim: Pooled low-resource training corpora for multilingual TTS contain significant quality variance that requires aggressive filtering before flow-matching models can be reliably trained. source: §4.2.2 evidence: WER-based filtering using IndicWhisper reduced the Hindi training set from 1.3M to 1M utterances; a partial recovery via CTC alignment scores added 30K back, indicating that roughly 23% of the raw data was too noisy or misaligned to be usable. confidence: medium relevance: low - claim_id: speaker_prompting_benefit_depends_on_phonology role: refines claim: The relative benefit of speaker-prompted duration conditioning over text-and-alignment conditioning is not uniform and depends on the phonological and prosodic characteristics of the target language. source: §5.1, §5.2, Table 3, Table 4, Table 5 evidence: Tamil, with complex phoneme structures and high prosodic variability, benefits consistently from speaker prompting across both objective WER and human SMOS; Marathi shows a clear regression; Hindi and Bengali exhibit mixed patterns depending on test subset. confidence: medium relevance: high limitations: - All experiments are conducted on speech infilling (50% sentence masking) rather than full-sentence TTS generation, so intelligibility and speaker-similarity results may not transfer directly to the standard single-utterance TTS evaluation setting. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.acl-long.1252 published_date: "2025-07-27" entry_date: '2026-07-27' year: 2025 venue: ACL task: - TTS - SCA architecture: - hybrid relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: - hybrid_generation claims: - claim_id: acoustic_accent_match_improves_chatbot_reception role: supports claim: Incorporating an acoustic accent that matches a user's linguistic community improves chatbot warmth, similarity, and engagement preference even when the underlying text remains in standard dialect. source: §4.4.2, Figure 4 evidence: The AA-accent + SAE-text spoken chatbot configuration outperforms the SAE baseline across all evaluated Likert dimensions (comprehension, communication ease, similarity, role appropriateness, engagement preference) among 8 AAE-speaking evaluators. confidence: medium relevance: high - claim_id: high_intensity_text_dialect_produces_caricature role: complicates claim: Text-based dialect generation at high intensity produces caricature-like representations that degrade perceived inoffensiveness and role appropriateness with speakers of that dialect. source: §4.4.1, Figure 3 evidence: High-AAE text chatbots receive the worst inoffensiveness scores among all 9 AAE text variants, with near-neutral or negative scores for trustworthiness and role appropriateness, which the authors attribute to caricature-like over-modification rather than authentic AAE. confidence: medium relevance: low - claim_id: phonetic_modification_dominates_llm_generated_aae role: supports claim: Phonetic modification dominates LLM-generated AAE output, while semantic feature generation remains sparse across all tested models and intensity levels. source: §4.3, Figure 2 evidence: Claude-Sonnet-3.5-based AAE feature tagging (91% accuracy) shows phonetic features dominate across all three LLMs' generated AAE text, while semantic AAE features remain sparse regardless of intensity level. confidence: medium relevance: low - claim_id: tts_trained_on_standard_dialect_degrades_on_minority_accent role: complicates claim: TTS models trained predominantly on standard-dialect data exhibit degraded speech clarity when reproducing minority accent features, limiting the fidelity of accent-conditioned spoken chatbots. source: §5 evidence: F5-TTS, trained predominantly on Standard American English data, shows decreasing speech clarity ratings as AAE dialect intensity increases in the spoken chatbot condition, even though it is applied only via a single CORAAL reference clip for accent. confidence: medium relevance: low limitations: - The evaluator pool is small (12 text, 8 spoken) from a single university demographic, limiting the statistical reliability of the 15-metric Likert evaluation framework. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.acl-long.313 published_date: "2025-07-27" entry_date: '2026-07-27' year: 2025 venue: ACL task: - TTS architecture: - flow-matching relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - flow_matching_generation claims: - claim_id: explicit_alignment_not_required_for_robust_naf_tts role: supports claim: Explicit phoneme-level alignment is not required for robust non-autoregressive zero-shot TTS; modality-specific representation refinement before feature concatenation can resolve text-speech alignment failures in alignment-free architectures. source: §3.2, §5.1 evidence: ConvNeXt V2 text refinement prior to concatenation resolves the alignment-failure mode observed in E2 TTS without reintroducing phoneme-level forced alignment or a duration predictor, achieving WER 2.42% on the newly released LibriSpeech-PC test-clean benchmark. confidence: high relevance: low - claim_id: entangled_semantic_acoustic_features_cause_alignment_failure role: complicates claim: Deep entanglement of semantic and acoustic features in a shared transformer produces systematic alignment failures in non-autoregressive TTS that persist across training scale and cannot be addressed by re-ranking. source: §3.2, §5.1, Table 4 evidence: Removing the ConvNeXt text-refinement branch (F5-TTS -Conv2Text) causes complete alignment failure, and applying the same refinement to E2 TTS's U-Net backbone instead degrades WER to 18.1%, showing the entanglement problem is backbone-specific and not resolved by re-ranking alone. confidence: high relevance: low - claim_id: sway_sampling_improves_faithfulness_without_retraining role: supports claim: Inference-time flow step scheduling that densifies early integration steps improves faithfulness and speaker similarity in flow-matching TTS without any retraining. source: §3.2, §5.2, Table 5 evidence: Sway Sampling improves WER from 2.84% to 2.41% and UTMOS from 3.70 to 3.89 at 32 NFE on LibriSpeech-PC test-clean without any retraining, and the same inference-time schedule also improves E2 TTS's WER from 2.95% to 2.84%, confirming portability across CFM-based TTS models. confidence: high relevance: high - claim_id: naf_flowmatching_matches_ar_at_100k_hour_scale role: supports claim: Non-autoregressive flow-matching TTS trained at 100K-hour scale can match autoregressive systems on intelligibility while achieving substantially lower real-time factor. source: §5, Table 1 evidence: At 16 NFE, F5-TTS achieves WER 2.53% with an RTF of 0.15, compared to Voicebox's RTF of 0.64 and DiTTo-TTS's 0.162, while remaining competitive with autoregressive systems on intelligibility across Seed-TTS test-en and test-zh. confidence: medium relevance: low - claim_id: missing_standardized_benchmarks_barrier_to_comparison role: supports claim: Absence of standardised benchmark subsets with released prompt lists is a structural barrier to fair cross-paper comparison in zero-shot TTS; community-released evaluation sets directly reduce this problem. source: §4 evidence: The paper notes that comparisons with VALL-E 2 and MELLE rely on scores from a different, incompatible LibriSpeech test-clean subset reported in prior work, motivating the release of the fixed-prompt-list LibriSpeech-PC benchmark to enable fair cross-paper comparison in zero-shot TTS. confidence: high relevance: low limitations: - Comparisons against VALL-E 2 and MELLE reuse numbers from a different LibriSpeech test-clean subset reported in prior work, which cannot be directly compared to the LibriSpeech-PC results introduced by this paper. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.acl-long.388 published_date: "2025-07-27" entry_date: '2026-07-27' year: 2025 venue: ACL task: - SCA architecture: - transformer-enc-dec relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - transformer_encoder_decoder claims: - claim_id: diva_distillation_without_instruction_responses role: supports claim: Cross-modal context distillation using ASR transcription pairs can transfer instruction-following capabilities to a speech encoder without annotated speech responses. source: §3.2, §5.1.2 evidence: DiVA trains on CommonVoice speech-transcript pairs while its frozen text LLM supplies the distillation target, avoiding annotated response labels. confidence: high relevance: high - claim_id: diva_distillation_avoids_capability_forgetting role: supports claim: Distillation-based speech-language-model training can avoid the instruction-following capability loss observed after narrow supervised fine-tuning. source: §5.1.2, Table 3 evidence: DiVA produced no instruction-ignoring outputs in the reported QA analysis, while Qwen Audio transcribed rather than answered 30% of questions. confidence: high relevance: low - claim_id: diva_joint_alignment_and_distillation_required role: supports claim: Input-token alignment and output-distribution distillation provide complementary supervision and are both required for stable cross-modal instruction transfer. source: §6, Table 5 evidence: Output distillation alone failed on translation and emotion tasks, whereas input alignment alone yielded mostly incoherent generations. confidence: high relevance: high - claim_id: diva_whisper_qformer_initialization role: supports claim: Initialising cross-attention from a pretrained speech decoder provides a transferable audio-to-text alignment bias for spoken question answering. source: §6, Table 5 evidence: Random Q-Former initialization reduced QA performance by 13.7% relative to initialization from Whisper decoder cross-attention weights. confidence: high relevance: low - claim_id: diva_human_preference_diverges_from_automatic_scores role: complicates claim: Human preference can diverge materially from automatic task benchmarks when evaluating spoken assistants. source: §5.2 evidence: Across 522 judgments from 53 Prolific participants, DiVA won 72% of pairwise preferences against Qwen 2 Audio despite trailing it on several automatic benchmarks. confidence: high relevance: high limitations: - The preference test used crowd workers rather than domain experts, and its prompts were not released. - Evaluation was single-turn, so the preference result does not establish multi-turn conversational quality. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.acl-long.598 published_date: "2025-07-27" entry_date: '2026-07-27' year: 2025 venue: ACL task: - TTS architecture: - autoregressive-LM - flow-matching - hybrid relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: - autoregressive_generation - flow_matching_generation - hybrid_generation claims: - claim_id: cross_model_preference_pairs_yield_larger_gains role: supports claim: Preference alignment with cross-model preference pairs yields larger intelligibility improvements than single-model self-comparison pairs for TTS post-training. source: §5.2, Table 5 evidence: Ablation shows that combining intra-model and inter-model (cross-model) preference pairs yields larger WER reductions than intra-model self-comparison pairs alone, with inter-pairs and intra-pairs shown to be complementary. confidence: medium relevance: high - claim_id: dpo_derivable_for_flow_matching_and_masked_generative role: supports claim: DPO objectives can be derived for flow-matching and masked generative TTS architectures by operating in velocity space and masked token distributions, respectively, without requiring explicit reward models. source: §4.2, §4.3 evidence: The paper derives DPO-FM by showing the log-ratio of learned versus reference velocity fields at each flow timestep is equivalent to an MSE difference in predicted velocities, and derives DPO-MGM operating on masked token distributions, enabling principled preference alignment for F5-TTS and MaskGCT respectively. confidence: high relevance: high - claim_id: intelligibility_gains_generalize_to_unseen_languages role: supports claim: Intelligibility gains from preference alignment on English and Chinese data generalize to unseen languages (Japanese, Korean, German, French) without degrading speaker similarity. source: §5.1, Table 10 evidence: INTP alignment trained only on English and Chinese data reduces WER substantially when evaluated on Japanese, Korean, German, and French via GTSinger, without measurable loss in speaker similarity. confidence: high relevance: high - claim_id: weak_to_strong_generalization_in_speech_alignment role: supports claim: Preference datasets constructed from weaker TTS generators effectively improve stronger models outside the training distribution, demonstrating weak-to-strong generalization in speech alignment. source: §5.1, Table 4 evidence: Preference data built from the weaker ARS, F5-TTS, and MaskGCT generators improves the stronger CosyVoice 2 (average WER 13.09 to 10.56) and the authors' own Ints model, which were not part of INTP's construction. confidence: high relevance: high - claim_id: iterative_alignment_diminishing_returns role: refines claim: Iterative preference alignment yields diminishing returns across rounds, suggesting that base model capability is the primary ceiling on alignment effectiveness. source: §5.3, Table 6 evidence: Ints v1 (INTP-aligned) generates new preference data to train Ints v2, reducing average WER from 7.96 to 7.10, a smaller relative gain than the initial alignment round, suggesting base model capability bounds further iterative improvement. confidence: medium relevance: high limitations: - All evaluation uses the SeedTTS test sets (English and Chinese); no evaluation on genuinely out-of-distribution natural speech (e.g., radio or phone-call recordings) is provided. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.acl-long.911 published_date: "2025-07-27" entry_date: '2026-07-27' year: 2025 venue: ACL task: - TTS - evaluation architecture: - transformer-enc-dec relevance: high evidence_role: - core_evidence current_role: active_evidence method_family: - transformer_encoder_decoder claims: - claim_id: contextual_prompts_produce_more_appropriate_speech role: supports claim: Incorporating conversational and environmental context into TTS prompts produces measurably more contextually appropriate speech than attribute-label prompts, even when the model architecture is otherwise unchanged. source: §4.3, Table 2 evidence: DNA-TTS, trained with situated dialogue/narrative/action context, achieves higher MOS-D and MOS-S scores (3.85, 3.83) than none-prompt TTS baselines (3.65-3.78) on the DNASpeech leaderboard. confidence: medium relevance: high - claim_id: codec_tts_outperforms_prompt_based_on_situated_benchmark role: supports claim: Codec-based TTS systems, despite not being designed for contextual conditioning, outperform purpose-built prompt-based baselines on most metrics when evaluated on a situated TTS benchmark. source: §4.4, Table 2 evidence: VoiceCraft, a codec TTS system not designed for contextual conditioning, leads MOS-D (3.88) and MOS-S (3.89) on the CS-TTS leaderboard, marginally ahead of the purpose-built DNA-TTS baseline (3.85 / 3.83). confidence: medium relevance: high - claim_id: dialogue_context_benefit_bounded_beyond_8_turns role: complicates claim: 'The benefit of dialogue context in speech generation is bounded: performance peaks at 4-6 dialogue turns and degrades significantly beyond 8 turns, suggesting that long-context integration remains an open problem.' source: Appendix F, Table 3 evidence: Analysis of dialogue turn count shows both DNA-TTS and VALL-E peak in performance at 4-6 turns of context and degrade as context grows beyond 8 turns. confidence: medium relevance: low - claim_id: movie_scripts_viable_source_for_contextual_tts_data role: supports claim: Movie scripts provide a viable automated source for constructing contextually rich TTS datasets, with speech quality comparable to purpose-built corpora on standard naturalness metrics. source: §3.4, Table 1 evidence: 'Models trained on the movie-derived DNASpeech dataset achieve MOS comparable to purpose-built corpora (FastSpeech2 on DNASpeech: 4.01 vs. 3.97 on LJSpeech), despite the dataset being constructed via an automated alignment and filtering pipeline rather than studio-recorded from scratch.' confidence: medium relevance: high limitations: - The human evaluator pool is small (8 individuals, Kendall's W >= 0.5 threshold), which may reduce the reliability of the four task-specific MOS variant estimates. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.acl-long.937 published_date: "2025-07-27" entry_date: '2026-07-27' year: 2025 venue: ACL task: - codec architecture: - VAE relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - variational_latent_generation claims: - claim_id: domain_adaptive_partitioning_within_a_shared_codebook_resolves role: supports claim: Domain-adaptive partitioning within a shared codebook resolves the performance degradation that plagues unified single-codebook codecs on mixed multi-domain audio. source: §3.2, §5.3, Tables 2, 4, 6 evidence: Removing the partitioned codebook raises Mel Distance from 0.344 to 0.487 on speech and from 0.396 to 0.506 on music in ablation; the full UniCodec outperforms WavTokenizer (unified) on all three domains in both objective and MUSHRA evaluation. confidence: high relevance: high - claim_id: self_supervised_mask_prediction_objectives_integrated_directly_into role: supports claim: Self-supervised mask prediction objectives integrated directly into codec training improve semantic representation without requiring external pretrained SSL encoders. source: §3.4, §5.2, Table 5 evidence: The semantic training stage improves RAVDESS classification accuracy from 36.81 to 40.28 and Audio-MNIST from 69.84 to 70.94 compared to the codec without this stage; no auxiliary SSL model is used. confidence: high relevance: low - claim_id: joint_training_of_reconstruction_and_self_supervised_semantic role: complicates claim: Joint training of reconstruction and self-supervised semantic objectives from scratch degrades single-codebook codec performance; sequential staging is required. source: §3.4 evidence: Preliminary experiments show that training reconstruction, mask prediction, and contrastive loss simultaneously is infeasible; the two-stage approach (acoustic training first, then semantic stage) is necessary for stable convergence. confidence: high relevance: low - claim_id: large_scale_diverse_audio_training_data_introduces_noise role: complicates claim: Large-scale diverse audio training data introduces noise that degrades codec reconstruction quality, requiring targeted fine-tuning on curated high-quality subsets. source: §4, §5.3, Appendix C, Table 6 evidence: Training on 80K hours of mixed data without the fine-tuning stage raises Mel Distance by 0.103 on speech (0.448 vs 0.344) relative to the fine-tuned model; high-quality fine-tuning on LibriTTS clean, VCTK, and LJSpeech recovers most of the degradation. confidence: high relevance: high - claim_id: single_codebook_codecs_at_75_tokens_per_second role: supports claim: Single-codebook codecs at 75 tokens per second can achieve acoustic reconstruction quality competitive with multi-layer RVQ codecs operating at 600 tokens per second when combined with domain-adaptive design. source: §5.1, Table 3 evidence: UniCodec PESQ 3.03 and STOI 0.949 on LibriTTS test-clean exceeds Encodec (2.72, 0.939) and SpeechTokenizer (2.61, 0.917) at 600 TPS, with eight times fewer tokens per second. confidence: high relevance: low limitations: - 'Training is sensitive to noisy or low-quality input: large-scale noisy data alone degrades reconstruction quality, and the fine-tuning stage is essential. This dependency on curated data limits scalability.' - The single-codebook codec struggles to balance both acoustic reconstruction fidelity and semantic density across diverse domains simultaneously, as these objectives conflict. Streaming performance degrades relative to non-streaming inference. The paper does not demonstrate integration with an actual audio language model, leaving downstream generation quality unvalidated. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.acl-long.997 published_date: "2025-07-27" entry_date: '2026-07-27' year: 2025 venue: ACL task: - SCA architecture: - autoregressive-LM relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: - autoregressive_generation claims: - claim_id: alignslm_preference_opt_improves_semantic_coherence role: supports claim: Preference optimization with LLM-based semantic feedback improves long-range semantic coherence in textless spoken language models beyond next-token prediction training. source: §5.2, Table 2 evidence: DPO with Mistral-score preference data raises T-StoryCloze from 69.7% to 74.2% (1.3B model) and from 75.4% to 83.8% (7B model) over the pre-trained TWIST baseline; curriculum learning further improves both to 76.1% and 85.6% respectively. confidence: high relevance: high - claim_id: alignslm_llm_scoring_beats_perplexity role: supports claim: LLM-based semantic scoring provides more effective alignment signal for speech continuation than perplexity-based preference selection. source: §5.1, Table 1 evidence: On the 1.3B model, Align-SLM with Mistral score improves T-StoryCloze by +4.5 over the pre-trained baseline, whereas the PPL-based variant degrades T-StoryCloze by 2.0. confidence: high relevance: high - claim_id: alignslm_curriculum_learning_progressive_gains role: supports claim: Curriculum learning with iteratively tightened preference thresholds yields progressive semantic improvements in preference-optimized spoken language models. source: §5.3, Appendix E, Table 2 evidence: Two curriculum learning iterations on the 7B model improve T-StoryCloze from 83.8% (Align-SLM) to 85.6% (Align-SLM+CL) and GPT4-o score from 3.50 to 3.56. confidence: medium relevance: high - claim_id: alignslm_gains_not_uniform_grammar role: complicates claim: Semantic gains from preference optimization do not transfer uniformly to lexical and grammatical aspects of spoken language modeling. source: §5.5, Table 2 evidence: Align-SLM shows only marginal sBLIMP improvement (+1.3 for the 1.3B Mistral variant) while models with no preference training but better token designs (AudioLM, SyllableLM) outperform on grammatical correctness. confidence: medium relevance: high - claim_id: alignslm_gap_to_cascaded_topline role: complicates claim: Textless end-to-end spoken language models remain substantially behind cascaded ASR+LLM pipelines in semantic understanding even with preference optimization. source: §5.5, Table 2 evidence: The best Align-SLM variant achieves 77.9% sWUGGY and 86.8% T-StoryCloze, compared to 79.2% and 94.8% for the ASR+LLM cascade topline, a gap that preference optimization alone does not close. confidence: high relevance: high limitations: - Training and evaluation are restricted to English audiobook speech (LibriSpeech, MLS); generalisation to spontaneous, noisy, or multi-speaker speech is untested. - The paper does not compare against text-injecting SLMs (SPIRITLM, Moshi, VoxtLM) under matched data and compute budgets. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.findings-acl.1226 published_date: "2025-07-27" entry_date: '2026-07-27' year: 2025 venue: ACL task: - TTS architecture: - hybrid relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - hybrid_generation claims: - claim_id: multiagent_llm_produces_more_informative_scripts role: supports claim: Parallel multi-agent LLM architectures can produce more informative and engaging conversational scripts than single-prompt generation for long-form spoken content. source: §5.1, §5.3, Table 1, Table 2 evidence: Host-Guest-Writer system outperforms direct GPT-4 baseline on all five quantitative metrics and all five LLM-as-a-Judge dimensions across four topic categories; the single-agent ablation also shows substantially lower scores on all metrics. confidence: high relevance: low - claim_id: llm_predicted_style_improves_speech_expressiveness role: supports claim: LLM-predicted speaking style instructions improve speech expressiveness in zero-shot TTS when used for long-form conversational audio. source: §5.2, Figure 7 evidence: All CMOS scores (0.2-0.9) are positive and preference tests favour LLM-instructed synthesis over uninstructed synthesis across all four topic categories in a 9-evaluator study. confidence: medium relevance: high - claim_id: automatic_voice_role_matching_cannot_guarantee_alignment role: complicates claim: Automatic voice-role matching from a speaker pool cannot guarantee perceptual alignment between voice characteristics and speaker roles in conversational audio generation. source: §5.2, §7 evidence: 87.4% average pass rate implies roughly 1 in 8 voice assignments is judged inappropriate by participants; the voice pool is also limited to read-speech data (LibriTTS-R), and the paper notes the need for more conversational-style voices. confidence: medium relevance: high - claim_id: more_participants_not_monotonically_better role: supports claim: Increasing the number of participants in multi-agent dialogue generation does not monotonically improve content quality, and smaller groups with more focused interactions outperform larger ensembles. source: §5.3, Table 2 evidence: Ablation across 1-5 guest configurations shows that 2-guest setups score highest on Distinct-N, MATTR, and Semantic-Div; quality degrades progressively from 3 to 5 guests. confidence: medium relevance: low limitations: - Content evaluation uses GPT-4-as-a-Judge to evaluate GPT-4-generated outputs, introducing well-documented self-preference and position biases that the paper only partially mitigates by averaging comparison orders. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.findings-acl.470 published_date: "2025-07-27" entry_date: '2026-07-27' year: 2025 venue: ACL task: - SCA - evaluation architecture: [] relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: [] claims: - claim_id: text_llm_speech_extension_degrades_context_recall role: supports claim: Extending a pre-trained text LLM to process speech input substantially degrades its ability to recall and utilize long-range conversational context. source: §4.1, Table 2 evidence: All four evaluated open-source voice interaction models score 2.00-2.83 on ContextDialog (5-point GPT Score), compared to 4.03-4.10 for their text-only backbone LLMs; the gap is largest when the LLM backbone is frozen during speech adaptation. confidence: high relevance: low - claim_id: modality_specific_recall_asymmetry role: supports claim: Voice interaction models exhibit a modality-specific recall asymmetry, retrieving information from model-generated text responses more reliably than from user-spoken utterances. source: §4.1, Table 2, Figure 3 evidence: All four speech models score significantly higher on questions about the model's own past utterances than past user utterances (p < 0.01); attention maps confirm systematically lower attention weight on user-spoken supporting utterances. confidence: high relevance: low - claim_id: rag_no_consistent_improvement_context_recall role: complicates claim: Retrieval-augmented generation provides a practical route to improving multi-turn context recall in spoken dialogue models. source: §4.2, Figure 4, Table 3, Table 5 evidence: Augmenting four open-source voice interaction models with a top-k text retriever (e5-large-v2) yields no consistent improvement and often degrades performance; top-1 retrieval recall is only 0.58 and irrelevant retrieved utterances lower all models below baseline. confidence: medium relevance: low - claim_id: opensource_voice_models_underdeveloped_longcontext role: supports claim: Open-source voice interaction models show critically underdeveloped long-context capabilities compared to closed-source systems, even on simple factual recall tasks. source: §A.1.4, Table 7 evidence: gpt-4o-mini-audio-preview scores 3.83-4.50 overall, more than 1 point above the best open-source model (Lyra, 2.83), even though the closed-source API does not support a proper multi-turn spoken dialogue context. confidence: medium relevance: low limitations: - 'ContextDialog is entirely synthetic: questions are generated by GPT-4o and corresponding audio is synthesized by Fish Speech, which may cause the benchmark to underestimate recall difficulty on real user speech with natural disfluency and acoustic variability.' caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.findings-acl.534 published_date: "2025-07-27" entry_date: '2026-07-27' year: 2025 venue: ACL task: - SCA architecture: [] relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: [] claims: - claim_id: semantic_similarity_reliable_proxy_for_tts_data_filtering role: supports claim: Semantic similarity in embedding space can serve as a reliable proxy for linguistic equivalence when filtering TTS-synthesized speech instruction data, avoiding the over-rejection caused by WER-based criteria. source: §5.2, §5.3, Table 1 evidence: The SIM metric achieves over 98% consistency with WER when strict linguistic equivalence holds, and above 92.5% across all conditions and three TTS models on seven QA datasets. confidence: high relevance: low - claim_id: llm_continued_targets_weaker_than_human_annotated role: supports claim: Training large speech language models on LLM-continued alignment targets produces systematically weaker instruction-following capability than training on human-annotated responses, regardless of speech synthesis quality. source: §5.3, Table 3 evidence: Models trained with LLM-continued answers achieve 52.54 average ROUGE-L vs. 60.5 with human-annotated golden answers at identical data quality thresholds; the gap persists across all ablation settings. confidence: high relevance: high - claim_id: parallel_multillm_rewriting_improves_tts_data_usability role: supports claim: Parallel query rewriting with multiple structurally diverse LLMs improves TTS data usability compared to any single-LLM rewriting approach, due to complementary performance across text types. source: §5.4, Table 2 evidence: Using Llama-3, Phi-3, and Qwen2 together raises average Pass rate by 2-3 percentage points over the best single-LLM baseline across MeloTTS and ParlerTTS-Large in the single-speaker setting. confidence: medium relevance: low - claim_id: wer_filtering_causes_inappropriate_data_rejection role: complicates claim: WER-based quality filtering of TTS-synthesized speech instruction data causes inappropriate data rejection due to incidental ASR errors, distorting the resulting dataset distribution. source: §3, §5.3, Table 13 evidence: Strict WER filtering rejects samples where ASR recognition errors occur despite semantic equivalence; SIM-based filtering recovers these samples while maintaining downstream task performance, confirmed by case analysis. confidence: medium relevance: low limitations: - LSLM evaluation covers only four reading comprehension tasks measured with ROUGE-L, which may not reflect instruction-following quality in conversational or multi-turn settings, and all experiments are English-only. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2507.20140' published_date: "2025-07-27" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - flow-matching relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - flow_matching_generation claims: - claim_id: unlearning_via_randomization_suppresses_speaker_identity role: supports claim: Machine unlearning via randomization-based training objectives can selectively suppress specific speaker identities in zero-shot TTS while preserving synthesis quality for retained speakers. source: §5.2, Table 1 evidence: TGU achieves SIM-F of 0.169 (within the measured cross-speaker similarity range of 0.02-0.17) and retain-set SPK-SIM of 0.631, a drop of only 2.8% from the original model's 0.649, with WER-F of 2.4% comparable to the original's 2.1%. confidence: high relevance: high - claim_id: data_exclusion_insufficient_for_voice_privacy role: complicates claim: Excluding target speakers from the fine-tuning dataset is insufficient for voice privacy protection in zero-shot TTS, because these models generalize at inference time to replicate unseen speakers via in-context learning. source: §5.2, Table 1 evidence: Exact Unlearning and Fine Tuning on the retain set yield SIM-F of 0.687 and 0.675 respectively, nearly identical to the original model's 0.708, confirming the model continues to clone forgotten speakers not present in the fine-tuning data. confidence: high relevance: low - claim_id: gradient_reversal_unlearning_degrades_intelligibility role: complicates claim: Gradient-reversal and KL-divergence-based unlearning methods degrade speech intelligibility rather than achieving genuine speaker forgetting in voice-conditioned generative models, due to entanglement between speaker style and linguistic content. source: §5.2, Table 1 evidence: Negative Gradient achieves WER-F of 5.0% and KL divergence maximization achieves WER-F of 47.2%, compared to 2.4% for TGU. confidence: high relevance: low - claim_id: speaker_similarity_alone_insufficient_to_verify_unlearning role: supports claim: Speaker similarity metrics alone are insufficient to verify effective machine unlearning in generative speech models, as consistent failure patterns can yield low similarity scores without achieving the randomness needed to resist voice reconstruction. source: §4.3, §5.2, Table 1 evidence: Negative Gradient and KL methods achieve low SIM-F scores (0.402 and 0.179) but spk-ZRF-F values of 0.842 and 0.810, at or below the original model's baseline of 0.846, revealing patterned outputs that could be reverse-engineered; TGU achieves both low SIM-F (0.169) and elevated spk-ZRF-F (0.871). confidence: high relevance: high - claim_id: pretrained_model_as_own_teacher_for_unlearning role: supports claim: A pre-trained generative model can serve as its own teacher for unlearning by providing diverse speaker-identity targets, eliminating the need for aligned cross-speaker audio pairs. source: §4.2 evidence: TGU generates per-step training targets by running VoiceBox conditioned only on text (no audio prompt), producing a different speaker identity at each Gaussian initialization; these unconditional outputs replace paired cross-speaker audio as forget-set targets. confidence: medium relevance: low limitations: - All experiments are conducted on a single TTS architecture (VoiceBox with mel-spectrogram representation); transferability of TGU to codec-based autoregressive ZS-TTS systems is untested. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2507.20731' published_date: "2025-07-28" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - GAN relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - gan_adversarial_generation claims: - claim_id: linear_degradation_structure_reduces_learning_burden role: supports claim: Explicit incorporation of known linear degradation structure (the mel filterbank) into vocoder design reduces the learning burden on the neural component and improves parameter efficiency. source: §4.3, Table 4 evidence: Removing the RND decomposition and replacing it with direct black-box mel-to-spectrum mapping degrades PESQ from 3.987 to 3.655 and causes near-total pitch tracking failure on LJSpeech. confidence: high relevance: low - claim_id: orthogonality_enforcement_important_for_tf_vocoders role: supports claim: Enforcing orthogonality between range-space and null-space components in T-F vocoder design is important; making the decomposition matrices learnable degrades reconstruction quality. source: §4.3, Table 4, Figure 6 evidence: Setting A and A† as learnable degrades PESQ from 3.987 to 3.645 and V/UV F1 from 0.9714 to 0.1164 on LJSpeech, indicating the learned matrices no longer maintain orthogonality. confidence: high relevance: low - claim_id: tf_vocoders_can_close_quality_gap_with_time_domain_gans role: supports claim: T-F domain vocoders can close the quality gap with large-scale time-domain GAN vocoders while remaining substantially more parameter-efficient and faster on CPU. source: §4.3, Tables 1, 2, 3 evidence: RNDVoC (3.14M parameters) achieves PESQ 4.226 and MUSHRA 80.74 on LibriTTS, matching or exceeding BigVGAN (112M parameters, PESQ 4.027, MUSHRA 79.33) trained for 1M steps, while running at nearly 10x real-time speed on CPU. confidence: high relevance: high - claim_id: naive_tf_vocoders_remain_inferior_without_subband_design role: complicates claim: The quality advantage of T-F domain vocoders over large-scale time-domain models depends on architectural choices for sub-band modeling; naive T-F vocoders remain inferior to well-tuned time-domain GAN vocoders at equivalent scale. source: §4.3, Table 1 evidence: Vocos (13.46M) achieves PESQ 3.522 on LJSpeech while BigVGAN (112M) reaches 4.107; only the proposed hierarchical sub-band design closes this gap at 3.14M parameters. confidence: medium relevance: low - claim_id: omnidirectional_phase_loss_improves_objective_metrics role: supports claim: Extending phase loss to capture differential relations across all eight adjacent T-F bins improves objective vocoder metrics relative to sparse two-direction anti-wrapping losses. source: §4.3, Table 4 evidence: Replacing the proposed omnidirectional phase loss with the prior anti-wrapping formulation reduces PESQ from 3.987 to 3.892 and MCD from 2.047 to 2.214 on LJSpeech. confidence: medium relevance: low limitations: - Evaluation is restricted to single-speaker (LJSpeech) and multi-speaker clean-speech (LibriTTS) benchmarks; no quantitative generalisation results are reported for singing, emotional speech, or noisy conditions. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.conll-1.9 published_date: "2025-07-31" entry_date: '2026-07-27' year: 2025 venue: CoNLL task: - TTS - evaluation architecture: [] relevance: medium evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: naturalness_ratings_do_not_fully_reflect_whether_tts role: complicates claim: Naturalness ratings do not fully reflect whether TTS prosody is linguistically appropriate. source: §5.3, Figure 3 evidence: Syntactic Sensitivity Score precision and recall diverge from the MOS ranking across Tacotron2, SpeechT5, and Parler-TTS; specifically, recall does not follow MOS order, revealing that missed boundary errors have less impact on perceived naturalness than spurious boundary errors — a dimension MOS does not capture independently. confidence: high relevance: high - claim_id: punctuation_marks_are_the_dominant_surface_cue_for role: supports claim: Punctuation marks are the dominant surface cue for intonational boundary placement in TTS systems, overriding syntactic and semantic information in ambiguous contexts. source: §4.4, §5.3 evidence: All three systems place boundaries reliably only when commas are present; without comma cues in garden path and attachment ambiguity sentences, systems default to the statistically most frequent pattern regardless of the intended structure, and semantic bias in prepositional phrase attachment has no measurable effect on boundary placement. confidence: high relevance: high - claim_id: transformer_based_tts_architectures_encode_partial_clause_boundary role: supports claim: Transformer-based TTS architectures encode partial clause boundary information for prosody that LSTM-based systems do not. source: §6.1, Figure 4 evidence: LASSO regression selects `is_clause_boundary` as a significant predictor of pause duration for SpeechT5 and Parler-TTS but not for Tacotron2, which relies more on token position and length features. confidence: high relevance: low - claim_id: increasing_tts_training_exposure_to_syntactically_diverse_speech role: refines claim: Increasing TTS training exposure to syntactically diverse speech can improve prosodic alignment with syntactic structure for structural cues, but not for semantic disambiguation cues. source: §7.3, Figure 5 evidence: Fine-tuning Parler-TTS on Jenny corpus prepositional phrase data produces more differentiated pause patterns correlated with syntactic function of ambiguous words; fine-tuning on synthetic semantic-bias data increases pause duration generally but does not teach the model to distinguish high from low attachment based on semantic context. confidence: high relevance: low limitations: - The study measures intonational boundaries through duration proxies only (pause insertion and pre-boundary lengthening), omitting pitch contour and intensity changes that also contribute to boundary perception. The ToBI framework identifies multiple boundary strength levels that this methodology cannot distinguish. - Evaluation is limited to English and to three specific TTS systems trained on audiobook speech, which under-represents prosodically disambiguating speech patterns relative to conversational speech. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.icnlsp-1.34 published_date: "2025-08-01" entry_date: '2026-07-27' year: 2025 venue: workshop task: - TTS architecture: - autoregressive-LM relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: - autoregressive_generation claims: - claim_id: preference_alignment_via_dpo_can_match_the_perceptual role: supports claim: Preference alignment via DPO can match the perceptual quality of supervised fine-tuning on human-labeled TTS data when preference pairs are constructed automatically from model-generated samples. source: §3.5, Table 1, Table 2, Table 5 evidence: 'On VCTK, DPO fine-tuning with LLM-augmented texts (V4) achieves WER 0.013, SS 0.482, and UTMOS 4.108 from the base xTTSv2 checkpoint, outperforming classical fine-tuning on human-labeled data (V1: WER 0.041, SS 0.500, UTMOS 3.685) on intelligibility and naturalness. CMOS evaluation confirms no statistically significant perceptual difference (all CM p > 0.30).' confidence: high relevance: high - claim_id: the_initialisation_point_before_dpo_fine_tuning_critically role: complicates claim: The initialisation point before DPO fine-tuning critically determines outcome quality, with checkpoint diversity being a more important factor than label quality of the preceding supervised stage. source: §3.6, Table 1, Table 2 evidence: DPO applied from the ASR-fine-tuned checkpoint (L5/V5) consistently underperforms DPO from the base model (L3/V3), with V5 reaching WER 0.273 versus V3's WER 0.014. The explanation offered is that standard fine-tuning narrows generation diversity, leaving insufficient variation for DPO preference signal. confidence: high relevance: high - claim_id: augmenting_unlabeled_tts_training_data_with_llm_generated role: supports claim: Augmenting unlabeled TTS training data with LLM-generated texts improves synthesis naturalness when the dataset's own texts are structurally mismatched with natural language. source: §3.5, Table 1, Table 2 evidence: On VCTK, which contains newspaper sentences well-aligned with LLM output, adding 15,000 LLM-generated texts (V4 vs. V3) raises UTMOS from 4.009 to 4.108. On LJSpeech, which contains audiobook fragments (partial sentences), the improvement is smaller and WER increases, attributed to domain mismatch. confidence: high relevance: high - claim_id: speaker_similarity_is_the_most_difficult_quality_dimension role: complicates claim: Speaker similarity is the most difficult quality dimension to preserve when fine-tuning TTS on ASR-transcribed or automatically constructed data, even when naturalness and intelligibility are maintained. source: §3.5, §3.7, Table 4, Table 5 evidence: Human-labeled fine-tuning (L1/V1) achieves the highest SS scores across both datasets (SS 0.481 on LJSpeech, 0.500 on VCTK). DPO from the base model recovers intelligibility and naturalness but trails on SS. Statistical significance testing shows SS degradation is significant for checkpoint-initialised DPO (Exp 6 vs. Exp 1/2, p < 0.001), while naturalness differences are not significant. confidence: high relevance: high limitations: - All experiments use xTTSv2 as the sole base model; it is unknown whether the same training recipe and checkpoint-initialisation finding generalise to other LM-based TTS architectures. Evaluation covers English only despite xTTSv2 being multilingual. - The comparison between fine-tuning from scratch on ASR-labeled versus human-labeled data is not performed; the human-labeled baseline starts from the same pre-trained xTTSv2 checkpoint, not a randomly initialised model, which limits conclusions about the true cost of label-free training. CMOS subgroup sample sizes (as few as 5 per accent in some comparisons) are too small for reliable accent-level conclusions. The bias implications of using ASR and LLM components for underrepresented speaker groups are acknowledged but not empirically quantified. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.sigdial-1.21 published_date: "2025-08-01" entry_date: '2026-07-27' year: 2025 venue: workshop task: - SCA architecture: - hybrid relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - hybrid_generation claims: - claim_id: acoustic_only_realtime_turntaking_inference_feasible role: supports claim: Purely acoustic, real-time turn-taking inference is feasible without relying on ASR or linguistic features. source: §5.3, Figure 6 evidence: TRPDformer achieves AUC = 0.83 on CEJC using only CPC audio representations and a causal transformer, outperforming a VAD baseline with no transcription step. confidence: high relevance: low - claim_id: vap_models_not_directly_transferable_to_trp_detection role: complicates claim: Voice activity projection models are not directly transferable to transition relevance point detection, even though the tasks appear structurally similar. source: §4.2 evidence: Statistical analysis of CEJC shows that a theoretically perfect VAP model would achieve at most recall = 0.60 and precision = 0.71 on TRP detection, below TRPDformer's empirical performance, because speaker shift and TRP are not equivalent events. confidence: high relevance: high - claim_id: learned_timing_models_perceived_more_natural_than_vad role: supports claim: Learned acoustic response-timing models produce perceptibly more natural dialogue turn-taking than fixed-threshold silence detection. source: §5.5, Figure 8 evidence: In a preference test with 40 raters per stimulus pair, subjects preferred TRPDformer's response timing over the VAD baseline, including in cases where both models misdetected TRPs. confidence: medium relevance: high - claim_id: framelevel_trp_models_may_not_generalize_across_languages role: complicates claim: Frame-level TRP models trained on spontaneous conversation may not generalise across languages or domains, limiting their direct applicability in multilingual or task-oriented dialogue systems. source: §6 evidence: TRPDformer is trained and evaluated solely on CEJC (Japanese everyday conversation); the paper identifies cross-lingual and cross-domain robustness as open future directions. confidence: low relevance: low limitations: - Evaluation is limited to a single language (Japanese) and a single corpus (CEJC); the paper does not report real-time processing latency or robustness to noise, both critical for deployment. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.sigdial-1.27 published_date: "2025-08-01" entry_date: '2026-07-27' year: 2025 venue: workshop task: - SCA - TTS architecture: - transformer-enc-dec relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - transformer_encoder_decoder claims: - claim_id: emotionconditioned_tts_in_sds_improves_emotion_appropriateness role: supports claim: Integrating emotion-conditioned TTS into a spoken dialogue system produces measurably more appropriate emotional speech than a neutral TTS baseline. source: §3.5, Table 2 evidence: Speech emotion appropriateness scores were 4.10 (proposed) vs. 1.70 (baseline) with a very large effect size (Cohen's d = 3.07, p < 0.001) across 20 human-judged conversations. confidence: medium relevance: high - claim_id: sentimentdriven_control_not_sufficient_for_engagement_gains role: complicates claim: Sentiment-driven emotional speech control does not straightforwardly translate into significantly higher user engagement in short task-oriented interactions. source: §3.5, Table 2 evidence: Engagement scores and number of SDS turns showed large effect sizes favouring the proposed system (d = 0.824, d = 0.831) but failed to reach statistical significance (p = 0.09, p = 0.07) with n = 20 conversations. confidence: low relevance: low - claim_id: discrete_contextfree_emotion_shifts_undermine_naturalness role: complicates claim: Discrete, context-free emotion label assignment produces step-change tonal shifts that may undermine conversational naturalness even when individual utterance emotion is appropriate. source: §3.5 evidence: Mean engagement remained around 3.0 on a 5-point scale, and the authors note that abrupt emotion transitions without considering prior context may have suppressed engagement gains. confidence: low relevance: high - claim_id: emotional_sds_needs_taskspecific_evaluation_metrics role: supports claim: Subjective evaluation of emotional spoken dialogue systems requires task-specific metrics beyond standard TTS naturalness; generic MOS-type ratings do not capture emotion appropriateness or conversational engagement. source: §3.3 evidence: The paper proposes a seven-item questionnaire covering RAG quality, task achievement, speech emotion appropriateness, and engagement, arguing existing evaluation frameworks lack appropriate metrics for social goals in emotional SDSs. confidence: medium relevance: high limitations: - The evaluation uses only 20 conversation samples (10 per system) conducted by the authors themselves, which is underpowered for the engagement and turn-count analyses and does not include independent users. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2508.00317' published_date: "2025-08-01" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: ssl_mos_finetuning_indomain_success_ood_gap role: supports claim: Fine-tuning self-supervised speech representations is sufficient to achieve high MOS prediction correlation in-domain, even with small labeled datasets, but this does not transfer reliably to out-of-domain evaluation settings. source: §II.A evidence: Fine-tuning SSL models for MOS prediction emerged as the dominant approach in VMC 2022, achieving high correlation even with very small amounts of labeled OOD data (down to 136 training samples), while VMC 2023 found no team achieved uniformly strong performance across all three tracks from a single model trained on shared data. confidence: medium relevance: high - claim_id: mos_predictors_fail_at_high_quality_ceiling role: complicates claim: Automatic MOS predictors consistently fail to discriminate among high-quality speech generation systems whose naturalness approaches human levels, indicating a ceiling effect in current SQA methodology. source: §II.C evidence: VMC 2024's zoomed-in track, evaluating high-quality systems near human naturalness, proved particularly difficult for automatic predictors, indicating current SQA methods struggle to discriminate among the best modern TTS systems. confidence: medium relevance: high - claim_id: community_challenges_accelerate_evaluation_progress role: supports claim: Community challenges with well-documented baselines and open datasets accelerate research progress in evaluation methodology more reliably than individual paper contributions. source: §I, §II.E evidence: AMC 2025 attracted 24 teams, the highest across all editions, and the paper argues via citation trend data that the VoiceMOS/AudioMOS Challenge series and open-source toolkits (MOSNet, DNSMOS, NISQA, SSL-MOS, UTMOS, TorchAudio-Squim, SHEET, VERSA) have been the primary drivers of progress in automatic speech quality assessment. confidence: low relevance: high - claim_id: quality_assessment_beyond_speech_needs_new_dimensions role: refines claim: Expanding automatic quality assessment beyond speech to music and general audio requires rethinking evaluation dimensions beyond MOS, as production quality, content enjoyment, and textual alignment reflect distinct perceptual dimensions that do not collapse to a single score. source: §II.D, §IV evidence: AMC 2025 introduced a text-to-music quality track and a multi-axis audio aesthetics evaluation track, with organisers noting that a single in-domain MOS-style score does not capture axes such as production quality, content enjoyment, and text alignment for music and general audio. confidence: medium relevance: high limitations: - The paper's central argument that challenges and open-source toolkits drive progress is supported by citation count trends rather than a controlled study; alternative explanations (larger models, more data, general SSL pre-training improvements) are not ruled out. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2508.01796' published_date: "2025-08-03" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - singing architecture: - diffusion - GAN relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - gan_adversarial_generation - diffusion_generation claims: - claim_id: mel_vocoders_underrepresent_highfreq_detectable_signature role: supports claim: Mel-spectrogram-to-waveform vocoders systematically underrepresent high-frequency spectral structure, producing a detectable signature that lightweight classifiers can exploit. source: §1, §4.2 evidence: A ConvNeXt classifier trained to distinguish real from synthetic spectrograms readily identifies baseline vocoder (Vocos, mdctGAN, NU-Wave 2) outputs as synthetic due to blurry, smoothed high-frequency energy patterns, whereas human observers rate baseline Vocos spectrograms as real only 24.3% of the time versus 78.1% for ground truth. confidence: medium relevance: high - claim_id: bandwidth_extension_helps_only_with_redesigned_vocoder role: refines claim: Inserting an explicit bandwidth extension step prior to vocoding can improve both spectrogram visual realism and perceived audio quality, but only when the vocoder is redesigned to accommodate the higher-resolution intermediate representation. source: §4.4, Figure 8 evidence: Adding the LSE bandwidth-extension step to the original (unmodified) Vocos vocoder slightly reduces audio quality (MOS 3.329), because the unmodified 1D-convolution vocoder struggles with the larger frequency resolution; the quality gain requires the Vocos2D redesign, which alone reaches MOS 4.157. confidence: medium relevance: high - claim_id: dit_adaptable_to_timefrequency_spectral_data role: supports claim: Diffusion models with transformer backbones can be adapted to time-frequency spectral data by restricting self-attention to the time axis and using per-frequency learned embeddings to handle the lack of spatial invariance across frequency. source: §3.2 evidence: The LSE model restricts DiT self-attention to the time axis only and handles inter-frequency communication through linear projections with per-frequency-bin learned embeddings, reconstructing a 592-bin full-bandwidth linear spectrogram from a mel spectrogram condition. confidence: medium relevance: low - claim_id: spectrogram_realism_and_audio_quality_partially_independent role: complicates claim: 'Spectrogram visual realism and waveform perceptual quality are partially independent dimensions: a system can improve on one while degrading on the other depending on how the vocoder processes the extended spectral representation.' source: §4.4 evidence: The Vocos2D-only ablation (without the LSE step) achieves a subjective MOS of 4.157, close to the full LSE+Vocos2D system's 4.176, suggesting most of the MOS gain comes from the vocoder redesign rather than the LSE step, even though LSE contributes materially to spectrogram-level visual realism. confidence: medium relevance: high limitations: - All training and evaluation data is assembled from a mix of privately recorded singing audio and public datasets under a custom pipeline; no standard public benchmark is used for final comparison. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2508.02013' published_date: "2025-08-04" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - SCA - evaluation architecture: [] relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: [] claims: - claim_id: srpa_quality_determined_by_text_lm_not_speech_architecture role: supports claim: The quality of spoken role-playing in both cascaded and end-to-end systems is more strongly determined by the underlying language model's text reasoning ability than by the speech synthesis or speech-to-speech architecture. source: §5.2 evidence: Cascaded systems built on strong text LMs, particularly Llama-3.1-8B, achieve the best overall English scores (0.957), outperforming GPT-4o Audio overall (0.921), while open-source end-to-end models (Qwen2.5-Omni, LLaMA-Omni) fall well below cascaded baselines across all dimensions. confidence: medium relevance: low - claim_id: finetuning_gains_larger_for_expressive_dims_than_interaction role: refines claim: End-to-end speech-language models exhibit larger gains from task-specific fine-tuning in expressive speech dimensions than in interaction or role fidelity dimensions, which remain text-LM bottlenecked. source: §5.2 evidence: 'Fine-tuning Qwen2.5-Omni on SpeechRole-Data substantially improves the resulting SpeechRole-Agent (0.801 English overall vs. 0.568 for the base model), with gains extending to unseen roles (OOD: 0.783), suggesting the benefit is not purely role memorisation.' confidence: medium relevance: low - claim_id: llm_pairwise_eval_correlates_with_human_judgement_multireference role: supports claim: LLM-based pairwise evaluation of speech quality can achieve high rank correlation with human expert judgements when using multi-reference normalisation, making automated SRPA benchmarking feasible at scale. source: §5.3, Table 3 evidence: Human-LLM agreement via Spearman's ρ ranges from 0.67 to 0.98 across metrics in English and 0.64 to 0.96 in Chinese, with Human-LLM Agreement rates exceeding 0.75 for most dimensions. confidence: high relevance: high - claim_id: opensource_e2e_lags_cascaded_on_prosody_emotion role: supports claim: Current open-source end-to-end speech models lag proprietary cascaded systems in prosody consistency and emotion appropriateness by margins large enough to constitute a capability gap, not merely a performance gap. source: §5.2, Table 2 evidence: GPT-4o Audio trails cascaded systems on prosody consistency (0.900 vs. 0.966 for Llama-3.1-8B) and emotion appropriateness (0.816 vs. 0.964), while open-source end-to-end Qwen2.5-Omni scores only 0.568 overall in English. confidence: medium relevance: low limitations: - All reference speech in SpeechRole-Eval is TTS-synthesised rather than recorded from human actors, meaning benchmark scores measure performance relative to synthetic reference quality rather than against a naturalness ceiling. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2507.20091' published_date: "2025-08-07" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - SCA - TTS architecture: - autoregressive-LM relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - autoregressive_generation claims: - claim_id: codec_token_speech_lms_trained_on_tens_of role: supports claim: Codec-token speech LMs trained on tens of thousands of hours of data do not develop prosodic understanding or generation as emergent capabilities through pre-training alone. source: §1, §4.2–4.4 evidence: A sentence-level [Global] token capturing aggregate prosodic extremity is optionally prepended to control expressive range during generation; the authors find it beneficial for generation tasks and detrimental for understanding tasks. confidence: high relevance: low - claim_id: replacing_codec_tokens_with_explicit_interpretable_word_level role: supports claim: Replacing codec tokens with explicit, interpretable word-level prosody annotations enables a pre-trained text LLM to acquire prosody processing capabilities — including contrastive focus, emotion recognition, and style transfer — without task-specific fine-tuning. source: §3.1, §4.2–4.4 evidence: The decoder adapts StyleTTS2's duration predictor and prosody predictor — replacing their dependence on StyleTTS2's diffusion-based style embedding with conditioning on the word-level values generated by the LLM — while leaving all other StyleTTS2 modules frozen at their pre-trained weights. confidence: high relevance: high - claim_id: disentangling_prosodic_and_linguistic_content_in_the_token role: supports claim: Disentangling prosodic and linguistic content in the token sequence trades voice quality fidelity for prosodic expressiveness and controllability. source: §3.1, §5 evidence: A sentence-level [Global] token capturing aggregate prosodic extremity is optionally prepended to control expressive range during generation; the authors find it beneficial for generation tasks and detrimental for understanding tasks. confidence: high relevance: low - claim_id: prosody_to_content_dependencies_detecting_emphasis_or_emotion role: supports claim: Prosody-to-content dependencies (detecting emphasis or emotion from speech) are more difficult for codec-token LMs than content-to-prosody generation. source: §4.3, Table 4 evidence: On prosody understanding tasks, ProsodyLM produces a statistically significant 6.6% log-probability increase for emphasised words (emphasis detection on EmphAssess), while both codec-token baselines show near-zero signal. confidence: high relevance: low - claim_id: word_level_prosody_tokens_impose_only_marginal_degradation role: supports claim: Word-level prosody tokens impose only marginal degradation on text language modelling capability when interleaved with transcription tokens in a joint sequence. source: §4.5, Table 9 evidence: Text perplexity on held-out audiobooks increases by only 2 points relative to a text-only fine-tuned Llama (11.70 vs. 13.87), indicating that prosody tokens add minimal interference to language modelling. confidence: high relevance: low limitations: - ProsodyLM is trained and evaluated exclusively on audiobooks (Librilight). Audiobook prosody is stylised and read-speech in character; generalisation to spontaneous conversation, emotional dialogue, or cross-domain settings is untested and likely requires substantial data re-collection or domain adaptation. - The decoder is frozen at StyleTTS2 pre-trained weights, which constrains output quality to that system's capability ceiling and limits the range of expressible voice characteristics. The five-dimensional prosody token cannot encode voice quality changes (breathiness, creakiness, vocal effort), which are important for fine-grained expressiveness. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2508.05385' published_date: "2025-08-07" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - transformer-enc-dec relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - transformer_encoder_decoder claims: - claim_id: automated_nv_annotation_matches_manual_annotation role: supports claim: Automated annotation pipelines for non-verbal vocalizations can match or exceed manually-annotated datasets in downstream NV generation and understanding tasks, while scaling at substantially lower cost. source: §5.1, Table 2; §5.2, Tables 4-5 evidence: F5-TTS fine-tuned on the automatically-annotated NVS(TSA) dataset achieves CLAP Scores of 0.187 (EN) and 0.179 (ZH), compared to 0.110/0.100 for the strongest comparable baseline trained on a smaller, more manually-curated dataset, representing relative gains of roughly 70-79%. confidence: medium relevance: high - claim_id: nv_tag_positional_accuracy_more_important_than_dataset_size role: supports claim: Positional accuracy of non-verbal tag annotations is a more important determinant of NV controllability in TTS than dataset size alone. source: §5.1, Table 2 evidence: The TSA (Temporal-Semantic Alignment) annotation variant consistently outperforms the simpler TBO (timestamp-based ordering) variant on the same underlying data, improving CLAP scores by roughly 0.07 and IMOS by roughly 0.2-0.3 in both English and Chinese. confidence: medium relevance: high - claim_id: framelevel_nv_detection_generalizes_crosslingual role: supports claim: Frame-level NV detection models trained exclusively on one language can generalise to structurally different languages without retraining, suggesting that acoustic features of non-verbal sounds are largely language-agnostic. source: §2.2; §3 evidence: The frame-level NV detection model, trained on English data only, achieves approximately 91% F1 across three test splits and performs comparably on Chinese test sets without additional training. confidence: medium relevance: low - claim_id: rulebased_nv_augmentation_worse_than_natural_data role: complicates claim: Rule-based NV data augmentation produces measurably worse NV controllability in fine-tuned TTS systems compared to models trained on naturally-occurring vocalizations. source: §5.1, Table 2 evidence: TTS models fine-tuned on the naturally-occurring NVS dataset outperform comparable systems trained on rule-based augmented data (NV clips concatenated or overlaid onto clean speech), which produce unnatural timbre mismatches. confidence: low relevance: low limitations: - The detection model is trained and evaluated on simulated test data (augmented speech + NV clips), not on naturally-occurring NV events in the wild; whether the ~91% F1 reflects real-world performance is untested. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2508.06870' published_date: "2025-08-09" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - transformer-enc-dec - GAN relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - transformer_encoder_decoder - gan_adversarial_generation claims: - claim_id: standard_tts_pipeline_intelligible_for_lowresource_tonal_script role: supports claim: Standard neural TTS pipelines (acoustic model + neural vocoder) can produce intelligible speech for previously unsupported tonal scripts with as little as 40 minutes of single-speaker training data. source: §5.3.5, §5.4 evidence: The Tacotron 2 + HiFi-GAN pipeline, trained from scratch on 818 utterances (~40 minutes) of Meitei Mayek speech, achieves naturalness MOS 3.34, pronunciation accuracy MOS 3.51, and overall MOS 3.43, with evaluators reporting intelligible output in most cases. confidence: medium relevance: high - claim_id: missing_tone_modeling_degrades_prosody_in_tonal_languages role: complicates claim: The absence of explicit tone modelling in a Tacotron 2 pipeline leads to degraded prosody on complex utterances in tonal languages, even when the training corpus is small. source: §2.1, §5.4 evidence: Manipuri's tonal contrast (falling vs. level tone) is not explicitly modelled; the tone diacritic is treated as another character in the phoneme table, and evaluators reported degraded rhythm specifically on long, syntactically complex sentences. confidence: low relevance: low - claim_id: novel_script_phoneme_mapping_necessary_prerequisite role: supports claim: Developing a phoneme mapping between a novel script and an established phoneme set (such as ARPAbet) is a necessary and non-trivial prerequisite before standard TTS toolkits can be applied to indigenous writing systems. source: §3.1.4, §6.2 evidence: The paper constructs a manual lookup table mapping each of the 55 unique Meitei Mayek characters to its closest ARPAbet equivalent, a prerequisite step without which the standard Tacotron 2 phoneme pipeline could not be applied to this script. confidence: medium relevance: low limitations: - The system is trained on a single speaker with only 40 minutes of audio, and no comparison to any baseline is possible; MOS scores are self-contained and cannot be situated relative to other low-resource systems. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2508.07711' published_date: "2025-08-11" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - GAN relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - gan_adversarial_generation claims: - claim_id: gan_not_necessary_with_explicit_phase_prediction role: supports claim: Adversarial training is not necessary for high-quality mel-spectrogram vocoders that explicitly predict the phase spectrum; frequency-weighted phase supervision can substitute for discriminator-provided phase signal. source: §IV.A, Table I evidence: FreeGAN achieves MOS 4.12 versus BigVGAN's 4.14 (p=0.61) and APNet2's 4.13 (p=0.76) with no discriminator and a frequency-weighted anti-wrapping phase loss in its place. confidence: high relevance: high - claim_id: gan_necessity_contingent_on_explicit_phase_prediction role: refines claim: 'The necessity of GAN in neural vocoders is contingent on the presence of explicit phase prediction: removing adversarial training without explicit phase modelling results in a quality gap that targeted improvements do not fully close.' source: §IV.B, Table III evidence: For systems without explicit phase prediction (HiFi-GAN, iSTFTNet, Vocos), removing GAN causes a noticeable quality drop that the proposed strategies only partially recover, while explicit-phase systems (APNet2, FreeV) show no significant drop. confidence: high relevance: low - claim_id: gan_elimination_reduces_cost_without_quality_loss role: supports claim: Eliminating GAN discriminators from vocoder training substantially reduces wall-clock training time and storage requirements with no measurable subjective quality cost for explicit-phase architectures. source: §IV.A, Table I evidence: FreeGAN trains in 191 seconds per epoch versus 1937 for BigVGAN, and eliminates the 70.7M discriminator parameters used by HiFi-GAN, while matching MOS. confidence: medium relevance: high - claim_id: utmos_and_mos_can_disagree role: complicates claim: Automatic MOS predictors (UTMOS) and subjective MOS scores can rank vocoder systems differently, making UTMOS an unreliable sole proxy for naturalness evaluation. source: §IV.A, Table I evidence: BigVGAN scores higher UTMOS (3.97) than FreeGAN (3.92) despite the two being statistically indistinguishable on subjective MOS (4.14 vs. 4.12, p=0.61). confidence: high relevance: high limitations: - All experiments use VCTK at 16 kHz, a constrained multi-speaker English setup; generalisation of the vocoder-quality findings to 24/44.1 kHz targets or non-English speakers is untested. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2508.08715' published_date: "2025-08-12" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - autoregressive-LM - flow-matching - GAN relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - autoregressive_generation - gan_adversarial_generation - flow_matching_generation claims: - claim_id: llm_based_tts_architectures_when_fine_tuned_with role: supports claim: LLM-based TTS architectures, when fine-tuned with appropriate speaker conditioning, substantially outperform conventional non-autoregressive models on low-resource language naturalness as measured by subjective MOS. source: §4.2 evidence: 'The architecture inherits CosyVoice''s three-component pipeline: a transformer-based autoregressive LM that generates speech tokens conditioned on text and language identifiers (zh, ma, ta for Mandarin, Malay, Tamil), a pretrained flow-matching model that maps speech tokens to continuous acoustic features, and a pretrained HiFi-GAN vocoder for waveform synthesis.' confidence: high relevance: high - claim_id: child_specific_vocal_characteristics_can_be_transferred_to role: supports claim: Child-specific vocal characteristics can be transferred to multilingual TTS through x-vector speaker conditioning without architectural modifications to the base model. source: §2.2, §3.2 evidence: 'Child-friendliness is introduced through speaker conditioning: x-vector embeddings extracted from children''s speech in each of the three languages are incorporated as speaker representations, steering the model toward age-specific vocal characteristics including higher pitch and distinctive prosodic patterns.' confidence: high relevance: low - claim_id: foundation_tts_models_originally_trained_on_high_resource role: supports claim: Foundation TTS models originally trained on high-resource languages can be adapted to low-resource language targets with limited domain data while retaining high intelligibility. source: §3.1, §4.3, Table 1 evidence: MultiGen fine-tunes CosyVoice-300M, an existing multilingual zero-shot TTS model, for five epochs on domain-specific data covering three target languages. confidence: high relevance: low - claim_id: intelligibility_gains_from_modern_tts_architectures_in_low role: supports claim: Intelligibility gains from modern TTS architectures in low-resource settings are more consistently captured by ASR-based CER than by subjective listener ratings, particularly when native annotators are limited. source: §4.3, Table 1 evidence: The paper reports evidence that intelligibility gains from modern TTS architectures in low-resource settings are more consistently captured by ASR-based CER than by subjective listener ratings, particularly when native annotators are limited. confidence: high relevance: high limitations: - Evaluation uses only 30 speech samples per language rated by 10 listeners each. The small scale means reported MOS differences, while large, should be treated cautiously — and no significance testing distinguishes genuine quality gaps from listener variability. All data is proprietary and not released, preventing independent replication. - The Singaporean-accented Mandarin training set contains only 1,400 utterances from a single child speaker, raising questions about voice diversity and generalisation to other child speakers. The paper does not evaluate speaker similarity to the target child voices, making it difficult to assess how faithfully the age-appropriate characteristics are reproduced. Malay and Tamil models are not evaluated against CosyBase, so it is unclear whether fine-tuning adds value beyond what a capable pretrained model already provides for those languages. All evaluation is limited to Singapore's specific linguistic and cultural context; the approach's generalisability to other low-resource child-speech settings is undemonstrated. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2508.08957' published_date: "2025-08-12" entry_date: '2026-07-27' year: 2025 venue: ASRU task: - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: pairwise_ranking_objectives_improve_mos_prediction role: supports claim: Incorporating pairwise ranking objectives into MOS prediction training improves system-level ranking consistency with human judgments beyond what standard regression losses achieve. source: §II.A, Table I evidence: QAMRO achieves SRCC of 0.972 (musical impression) and 0.916 (textual alignment) on MusicEval, compared to 0.845 and 0.779 for the same-backbone regression-only baseline. confidence: high relevance: high - claim_id: adaptive_margins_outperform_fixed_margin_ranking role: supports claim: Adaptive margins that scale with perceptual score gaps outperform fixed-margin ranking losses for audio quality assessment. source: §II.A, Table I — w/o Adaptive Margin ablation evidence: The ablation removing the adaptive margin component shows a measurable drop in SRCC relative to the full QAMRO objective. confidence: high relevance: high - claim_id: quality_aware_weighting_consistently_beneficial role: supports claim: Quality-aware upweighting of high-scoring sample pairs is a consistently beneficial component in automatic perceptual quality evaluation. source: §IV.A, §IV.B evidence: Ablations on both the MusicEval/CLAP backbone and the AES-Natural/Audiobox-Aesthetics backbone show removing quality-aware weighting degrades SRCC on both. confidence: high relevance: high - claim_id: training_objective_change_improves_pretrained_backbone_evaluators role: supports claim: Automatic MOS predictors built on pre-trained audio-text representations can be meaningfully improved by training objective changes without modifying the backbone architecture. source: §II.B, Table II evidence: QAMRO improves SRCC over the Audiobox-Aesthetics baseline across all four AES-Natural dimensions (e.g. content usefulness improves by 0.052 SRCC) using the same pretrained backbone. confidence: high relevance: high limitations: - All training and evaluation is conducted on the AudioMOS Challenge 2025 closed-set data only; generalisation to unseen TTS or VC systems outside this distribution is untested. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2504.12867' published_date: "2025-08-13" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - autoregressive-LM - flow-matching relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - autoregressive_generation - flow_matching_generation claims: - claim_id: fine_grained_natural_language_emotion_descriptions_provide_richer role: complicates claim: Fine-grained natural language emotion descriptions provide richer control over expressive speech synthesis than coarse categorical labels, at the cost of requiring emotion-specific training data. source: §2.1, §6.1 evidence: 'Training is two-stage: a standard TTS pretraining phase on large synthetic neutral speech (VoiceAssistant, 3,234h English; Belle, 6,418h Chinese), followed by emotional fine-tuning on EmoVoice-DB, LAION Got Talent, and in-house Chinese data.' confidence: high relevance: low - claim_id: parallel_phoneme_token_prediction_as_a_secondary_output role: supports claim: Parallel phoneme token prediction as a secondary output head reduces intelligibility errors in LLM-based TTS, particularly on challenging inputs such as rare words and tongue twisters. source: §6.2.1, §6.2.2, Table 5, Table 6 evidence: EmoVoice-PP (Phoneme Boost variant) adds a parallel prediction head that simultaneously outputs phoneme tokens (~11Hz) alongside audio tokens (~17Hz). confidence: high relevance: low - claim_id: llm_pretraining_initialisation_meaningfully_benefits_emotion_controllable_tts role: supports claim: 'LLM pretraining initialisation meaningfully benefits emotion-controllable TTS: models without it show substantially higher word error rates and weaker emotion transfer.' source: §6.2.4, Table 8 evidence: On the Chinese Secap test set, EmoVoice-PP achieves the best WER (7.6), emotion similarity (0.7939), and recall (0.434) among evaluated models, outperforming CosyVoice2 and GPT-series models, which struggle with non-English prosody and timbre. confidence: high relevance: low - claim_id: automatic_emotion_similarity_metrics_e_g_emotion2vec_cosine role: supports claim: Automatic emotion similarity metrics (e.g. emotion2vec cosine similarity) correlate reasonably at the system level but poorly at the utterance level with human perceptual judgments, limiting their utility for fine-grained model comparison. source: §7, Table 10 evidence: On the English EmoVoice-DB test set (objective), EmoVoice(1.5B) achieves the highest emotion similarity (0.9118) and recall rate (0.424) among open systems, with WER of 2.62 and UTMOS of 4.35. confidence: high relevance: high - claim_id: multimodal_llms_are_not_yet_reliable_judges_of role: supports claim: Multimodal LLMs are not yet reliable judges of emotional speech quality, exhibiting both low correlation with human ratings and inter-run instability. source: §7, Table 10 evidence: 'The ablation on LLM initialization shows it is essential: removing pretrained Qwen2.5 weights raises WER from 2.73 to 6.16 for the base model, confirming that LLM language understanding transfers meaningfully to emotional TTS quality.' confidence: high relevance: high limitations: - 'The English model is trained and evaluated entirely on synthetic data generated by GPT-4o-audio. Both EmoVoice-DB (training) and the test set are GPT-4o-audio outputs, creating circularity: the model learns to mimic GPT-4o-audio''s synthesis style rather than natural human emotional speech. Generalisation to real human emotional recordings or to out-of-distribution TTS systems is undemonstrated.' - The emotion recall evaluation omits three of seven emotion categories (disgusted, fearful, surprised) due to low recognition accuracy from emotion2vec, which limits the scope of the emotional expressiveness claims. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2508.09702' published_date: "2025-08-13" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS - evaluation architecture: [] relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: [] claims: - claim_id: quality_mismatch_degrades_zeroshot_generation role: supports claim: Quality mismatch between clean training corpora and real-world inference prompts is a meaningful source of degradation in zero-shot speech generation systems. source: §1 evidence: The paper motivates M3PDB by observing that zero-shot TTS systems trained on clean studio recordings degrade against noisy, incomplete, or out-of-domain inference prompts. confidence: medium relevance: low - claim_id: crosslingual_proxy_prompts_can_outperform_native role: supports claim: Cross-lingual proxy prompts selected by linguistic proximity can produce lower synthesis error than native-language prompts of lower audio quality. source: §3.2, Table 3 evidence: Cross-lingual prompts selected by M3PDB for Italian synthesis via XTTS-v2 yield CER of 2.78%, compared to 4.32% for native Italian prompts and 5.21% for randomly selected French prompts. confidence: medium relevance: low - claim_id: multimodal_annotation_improves_age_gender_accuracy role: supports claim: Multi-modal annotation that fuses visual and acoustic signals yields substantially higher age and gender labelling accuracy than audio-only or text-only annotation strategies. source: §5.1, Table 2 evidence: The multi-agent annotation system reduces age MAE from 12.54 (SpeechCraft) to 6.22 years and raises gender accuracy from 80.69% to 91.22%. confidence: medium relevance: high - claim_id: curated_database_outperforms_generative_face_to_speech role: supports claim: A curated, high-quality audio database outperforms generative face-to-speech synthesis as a source of speaker prompts when audio quality is the binding constraint. source: §5.3, Table 4 evidence: M3PDB-retrieved audio achieves comparable speaker similarity to Imaginary Voice face-to-speech synthesis (15.14% vs. 16.18% SS) while substantially improving UTMOSv2 (2.69 vs. 2.10). confidence: medium relevance: high - claim_id: cascaded_retrieval_viable_for_latency_constraints role: supports claim: Cascaded, interruptible similarity retrieval is a viable approach to prompt selection under strict latency constraints without dedicated hardware. source: §4.2, §5.4, Table 5 evidence: The online two-stage selection strategy improves SS from 39.53% to 44.88% and CER from 4.21% to 1.17% over using degraded input directly, without specialized hardware. confidence: medium relevance: low limitations: - No subjective listening tests are reported anywhere in the paper; all quality metrics (UTMOSv2, automatic speaker similarity via cosine embeddings, Emotion2Vec emotion similarity) are automatic predictors. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2503.04721' published_date: "2025-08-16" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - SCA - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: explicit_turn_state_reduces_pause_interruptions role: supports claim: Explicit turn-state prediction modules reduce unwanted interruptions during intra-turn speaker pauses more effectively than purely end-to-end architectures trained on dialogue data. source: §V, Table III evidence: Freeze-Omni, which uses an explicit turn-state classifier, reduces Takeover Rate to 0.481-0.642 during pause handling versus 0.934-0.985 for end-to-end dGSLM and Moshi. confidence: high relevance: high - claim_id: e2e_lower_latency_but_less_coherent_on_interruption role: supports claim: End-to-end full-duplex speech models respond with lower latency but produce less semantically coherent replies when confronted with user interruptions than cascaded systems with a frozen LLM backbone. source: §V, Table III evidence: Moshi responds to interruptions in 0.257s but scores only 0.765 on GPT-4o coherence, while Freeze-Omni responds in 1.409s with a coherence score of 3.615. confidence: high relevance: high - claim_id: backchanneling_largely_unsolved role: supports claim: Backchannel generation remains a largely unsolved capability in current full-duplex spoken dialogue models, with most open-source systems producing near-zero backchannel frequency. source: §V, Table III evidence: Only dGSLM produces backchannel responses at a measurable rate (Freq = 0.015); Moshi and Freeze-Omni produce near-zero backchannel frequency. confidence: high relevance: high - claim_id: commercial_systems_better_calibrated_turn_taking role: supports claim: Commercial full-duplex systems exhibit better-calibrated turn-taking behaviour on naturalistic data than open-source research models, suggesting that VAD design and deployment-scale tuning contribute substantially to perceived conversational naturalness. source: §V, Table III evidence: Gemini Live achieves the lowest pause-handling TOR (0.255-0.310) and the best backchannel-timing JSD (0.896) among all four evaluated systems. confidence: medium relevance: high - claim_id: automatic_metrics_can_be_reproducible_and_modelagnostic role: supports claim: Automatic metrics for full-duplex dialogue can be made reproducible and model-agnostic without human raters, enabling rapid cross-system comparisons that corpus-level statistics and user studies cannot efficiently provide. source: §I, §III-A evidence: Full-Duplex-Bench derives all four evaluation dimensions (pause handling, backchanneling, smooth turn-taking, interruption management) from automatic, time-aligned ASR transcripts and scenario-controlled stimuli, requiring no human annotators in the evaluation loop itself. confidence: high relevance: high limitations: - The benchmark evaluates only four models (three open-source, one commercial API) and is restricted to English, so conclusions about which architectural choices drive observed behaviours are correspondingly narrow. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-0305 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - singing - VC architecture: - flow-matching - transformer-enc-dec relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - transformer_encoder_decoder - flow_matching_generation claims: - claim_id: flow_matching_beats_gan_decoder_svc role: supports claim: Flow matching decoders produce higher audio quality than GAN-based decoders in singing voice conversion when conditioning signal quality is held constant. source: §4.1, §4.2, Table 1, Table 2 evidence: Replacing NeuCoSVC's GAN-based FastSVC decoder with a CFM module improves MCD from 8.634 to 7.220 and MOS-Naturalness from 3.47 to 3.80 on OpenSinger. confidence: high relevance: high - claim_id: ssl_matching_insufficient_alone_for_timbre_similarity role: refines claim: SSL feature matching prevents timbre leakage in singing voice conversion but is insufficient on its own for high timbre similarity, because target timbre is distributed across the full reference utterance rather than captured by sparse nearest-neighbour retrieval. source: §4.2, Table 1, Table 2 evidence: The SSL-matching-only ablation achieves SSIM 0.709 versus 0.692 for NeuCoSVC (marginal improvement), while adding speaker embeddings with the full DCAM raises SSIM to 0.754. confidence: medium relevance: low - claim_id: crossattention_fusion_improves_timbre_similarity role: supports claim: Cross-attention fusion of speaker embeddings and melody features with shared content queries improves timbre similarity and audio coherence over simple feature concatenation in conditional singing voice conversion. source: §4.2, Table 2 evidence: Removing the DCAM module while retaining speaker embeddings drops SSIM from 0.754 to 0.710 and MCD from 7.220 to 8.129. confidence: high relevance: low - claim_id: oneshot_svc_evaluation_remains_narrow role: complicates claim: One-shot singing voice conversion evaluations remain narrow in scope, limiting the generalisability of reported gains. source: §3.1, §3.4, §5 evidence: Experiments use a single Chinese singing dataset (OpenSinger), 20 samples for subjective evaluation with 15 listeners, and four unseen target speakers; cross-language, multi-domain, or noisy-environment generalisation is explicitly deferred to future work. confidence: medium relevance: high limitations: - The system is trained and evaluated exclusively on high-quality Chinese singing recorded in a professional studio; performance in noisy environments and cross-language settings is untested. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-0383 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - VC architecture: - transformer-enc-dec relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: - transformer_encoder_decoder claims: - claim_id: likabilityvc_discrete_unit_scalar_attribute_control role: supports claim: Discrete-unit voice conversion can be extended to control subjective perceptual attributes beyond speaker identity by adding a scalar conditioning signal to the synthesis model. source: §3, §5.3, Figure 6 evidence: A FastSpeech 2 model conditioned on HuBERT-based discrete units, ECAPA-TDNN speaker embeddings, and a target likability scalar successfully steered perceived likability for 3 of 4 speakers in pairwise preference tests. confidence: medium relevance: high - claim_id: likabilityvc_automatic_predictor_sufficient_proxy role: supports claim: Automatic likability predictors based on TDNN regression on crowd-sourced ratings can provide sufficient proxy labels to train large-scale perceptual attribute control systems. source: §4.2, Table 2 evidence: The predictor achieved LCC 0.46 and SRCC 0.49 with human ratings (p < 3e-17) and 74% binary classification accuracy; its outputs were used to automatically annotate the JVS and JTES corpora for VC training. confidence: medium relevance: high - claim_id: likabilityvc_control_identity_tension role: complicates claim: Strong likability control and speaker identity preservation are in tension in discrete-unit VC systems, particularly at extreme target values. source: §5.2, §5.3, Figures 4-6 evidence: At target likability = 2 (outside the training range), CER increased substantially for female speakers, and speaker m49 showed degraded speaker similarity and an unexpected subjective likability drop; the inference-time scalar multiplier partially mitigates but does not eliminate this trade-off. confidence: high relevance: high - claim_id: likabilityvc_per_speaker_failure role: complicates claim: Voice likability control demonstrates effective behaviour for majority speaker groups but can fail for individual speakers due to identity-likability interaction effects. source: §5.3, Figure 6 evidence: Three of four speakers showed significant preference differences between target -1 and 1; speaker m49 exhibited the opposite trend, attributed to failure to preserve speaker identity. confidence: medium relevance: high - claim_id: likabilityvc_multifactorial_demographic_modeling role: refines claim: Perceived voice likability is a multi-factorial attribute requiring demographic-stratified modelling, not a single group-level score. source: §2, §4.2, Table 2 evidence: The predictor uses four separate listener-group outputs (by gender and age); per-group LCC ranges from 0.36 to 0.41 while the aggregate LCC is 0.46, confirming systematic variation across listener demographics. confidence: medium relevance: high limitations: - The evaluation is limited to four speakers and one language (Japanese); the practical control range of the system is narrow, with predicted likability shifting only from approximately -0.51 to -0.23 despite targets spanning -2 to 2. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-0401 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: standard_mos_mushra_insufficient_for_modern_tts role: complicates claim: Standard MOS and MUSHRA protocols are insufficient for evaluating modern high-quality synthetic speech, and evaluation protocol saturation now drives the need for purpose-designed alternatives. source: §1, §2 evidence: The paper argues, building on a decade of meta-analyses (Wester et al. 2015, Kirkland et al. 2023), that as synthetic speech quality has risen, standard protocols increasingly fail to discriminate between systems without careful, purpose-designed adaptation. confidence: medium relevance: high - claim_id: missing_methodological_details_undermine_comparisons role: complicates claim: Missing methodological details in published evaluation reports, particularly the exact question posed to listeners and cohort composition, materially affect the validity of cross-paper comparisons. source: §1 evidence: The paper cites a decade of documented practice where the precise listener question, cohort composition, and protocol rationale are routinely omitted from published TTS evaluation reports, making replication and cross-paper comparison unreliable. confidence: high relevance: high - claim_id: shareable_recipe_schema_enables_replication role: supports claim: Standardising evaluation into a shareable recipe schema separates the evaluation design from its execution platform, enabling replication without re-implementing the full setup. source: §3.3 evidence: The replikant platform packages procedure, anonymised data, and configuration into a self-contained recipe bundle; several published evaluations (North Sami TTS, video summarisation) already reuse replikant-based recipes, demonstrating real-world replication. confidence: medium relevance: low - claim_id: review_process_can_improve_evaluation_quality role: supports claim: Conference review processes can meaningfully improve evaluation quality by assessing protocol justification, material appropriateness, and alignment between the statistical analysis and the evaluated hypothesis. source: §4, §4.1, §4.2 evidence: The paper proposes a checklist-style reviewer guideline grounded in ITU-T P.800, P.85, and prior meta-analyses, covering evaluation campaign design, material description, and analysis appropriateness. confidence: low relevance: low limitations: - Without formal adoption by Interspeech, ICASSP, or SSW programme committees, the reviewer guidelines remain advisory, and the paper does not evaluate whether recipe sharing actually increases replication attempts in practice. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-0406 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS architecture: - GAN relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - gan_adversarial_generation claims: - claim_id: pretrained_vocoder_refines_warped_speech_to_binaural role: supports claim: Pretrained monaural denoising vocoders can refine geometrically warped speech into perceptually natural binaural audio without exposure to any binaural training data. source: §3.3, §4.4 evidence: ZeroBAS applies a pretrained WaveFit vocoder iteratively to geometrically warped and amplitude-scaled mono audio, achieving MOS 4.07 on BSD versus 4.01 for the supervised BinauralGrad baseline. confidence: high relevance: high - claim_id: supervised_binaural_models_degrade_ood role: supports claim: Supervised mono-to-binaural synthesis models trained on small room-specific datasets degrade significantly under out-of-distribution acoustic conditions, including different rooms and languages. source: §4.5, Table 2 evidence: On the out-of-distribution TUT Mono-to-Binaural dataset (different rooms, French speakers), supervised baselines WarpNet and BinauralGrad drop to MOS 3.60 and 3.27 respectively, with BinauralGrad producing Gaussian noise artifacts on OOD samples. confidence: high relevance: high - claim_id: gtw_and_as_both_necessary_for_zeroshot_binaural role: supports claim: Geometric interaural time delay warping and amplitude scaling based on the inverse-square law are both necessary components for zero-shot binaural synthesis; removing either one substantially degrades perceptual quality. source: §5, Table 3 evidence: Removing GTW degrades MOS to 3.64 and removing AS degrades MOS to 2.93, versus 4.07 for the full ZeroBAS pipeline. confidence: high relevance: high - claim_id: in_distribution_benchmarks_insufficient_for_binaural_generalisation role: complicates claim: Standard in-distribution benchmarks for binaural synthesis are insufficient to assess generalisation; evaluating on held-out room conditions reveals large performance gaps that in-distribution results conceal. source: §4.2, §4.5 evidence: Supervised methods that are competitive with ZeroBAS on the in-distribution BSD dataset (differences not statistically significant by MUSHRA) fall significantly behind on the newly introduced out-of-distribution TUT Mono-to-Binaural dataset. confidence: high relevance: high limitations: - The method does not model room impulse responses or listener head shape; HRTF is implicit and generic rather than personalized. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-0408 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - SCA - TTS architecture: - transformer-enc-dec - GAN relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: - transformer_encoder_decoder - gan_adversarial_generation claims: - claim_id: combining_paralinguistic_intimacy_adaptation_with_linguistic_register_change role: supports claim: Combining paralinguistic intimacy adaptation with linguistic register change improves user satisfaction more than linguistic-only adaptation in multi-session dialogue. source: §5.2, Table 2 evidence: 'Paralinguistic adaptation (new contribution): A Tacotron 2-based TTS system, augmented with an intimacy embedding network replacing the speaker embedding, is trained on the SMOC Japanese spontaneous multimodal corpus.' confidence: high relevance: low - claim_id: intimacy_conditioned_tts_trained_on_spontaneous_conversational_data role: supports claim: Intimacy-conditioned TTS trained on spontaneous conversational data can reproduce prosodic features (speech rate and F0) that distinguish high- and low-intimacy speech. source: §5.1 evidence: 'Objective evaluation confirms the TTS model captures intimacy: speech rate RMSE and log-F0 RMSE are lower when synthesized speech intimacy level matches the ground-truth label.' confidence: high relevance: low - claim_id: binary_intimacy_labels_are_sufficient_to_produce_measurable role: supports claim: Binary intimacy labels are sufficient to produce measurable prosodic differences in synthesised speech, suggesting that coarse-grained conditioning captures style variation in spontaneous dialogue. source: §5.1 evidence: Binary intimacy labels switch in sync with linguistic style changes across the three experiment days. confidence: high relevance: low - claim_id: paralinguistic_intimacy_may_be_less_effectively_conveyed_through role: supports claim: Paralinguistic intimacy may be less effectively conveyed through speech prosody alone than through linguistic register, as gains in friendliness ratings remain modest even when prosodic style shifts. source: §5.2 evidence: Binary intimacy labels switch in sync with linguistic style changes across the three experiment days. confidence: high relevance: low limitations: - Only binary (high/low) intimacy labels are used; a continuous intimacy scale might enable finer-grained adaptation. The experiment spans only 3 days; longer interactions may reveal larger effects. The CG agent's gesture and emotion were frozen, which may have suppressed multimodal intimacy cues. The study is Japanese-language only, limiting generalizability. Future work intends to incorporate multimodal behaviors (gestures, facial expressions) and longer-term experiments. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-0433 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - VC architecture: - VAE relevance: high evidence_role: - core_evidence current_role: active_evidence method_family: - variational_latent_generation claims: - claim_id: h2nhvc_human_vc_needs_adaptation_for_nonhuman role: supports claim: Voice conversion architectures designed for human speech require non-trivial adaptation to generalise to non-human vocalizations with broad frequency ranges and transient-rich characteristics. source: §4.2.2, Table 1 evidence: Replacing the proposed preprocessing pipeline with a conventional speech-focused one degraded WER from 24.02% to 62.29% and MOS-S from 3.78 to 3.53. confidence: high relevance: high - claim_id: h2nhvc_style_isolation_reduces_leakage role: supports claim: Isolating style conditioning to the prior network and normalizing flow, and excluding it from the posterior encoder and decoder, reduces style leakage and improves speaker similarity in CVAE-based voice conversion. source: §4.2.3, Table 1 evidence: Adding the style embedding to the audio encoder and decoder (w/ SEED ablation) reduced MOS-S from 3.78 to 3.61, attributed to style overlap between the reference encoder output and latent acoustic tokens. confidence: high relevance: high - claim_id: h2nhvc_kl_annealing_mitigates_posterior_collapse role: supports claim: KL annealing mitigates posterior collapse in VAE-based voice conversion and improves linguistic content preservation, particularly for complex non-human vocalizations. source: §4.2.3, Table 1 evidence: Removing KL annealing increased CER from 15.48% to 28.89% and WER from 24.02% to 44.69%, while MOS scores changed minimally. confidence: high relevance: high - claim_id: h2nhvc_f0_estimation_fails_nonharmonic role: complicates claim: Existing fundamental frequency (F0) estimation methods are not reliable for sounds lacking a well-defined harmonic structure, constraining prosodic feature extraction in non-human voice conversion systems. source: §3.1 evidence: The authors tested frame-level F0 from non-human sounds but found existing estimators (Praat, CREPE, SPICE, PESTO) exhibited limitations due to absent harmonic structure; the system falls back to energy-only prosodic features. confidence: medium relevance: high limitations: - The dataset is entirely internal and evaluation uses only 9 human raters on an unspecified number of test samples; F0 estimation for non-human sounds remains unsolved, so the system omits pitch conditioning entirely. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-0438 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - VC architecture: - hybrid relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - hybrid_generation claims: - claim_id: linearvc_linear_transform_sufficient role: supports claim: Linear transformations of self-supervised speech features are sufficient for competitive voice conversion, without complex nonlinear decoders or model fine-tuning. source: §3.2, Table 1 evidence: LinearVC's single learned projection matrix on WavLM-Large layer 6 achieves WER 4.9%, EER 33.6%, and MUSHRA naturalness 62.5 on LibriSpeech test-clean, statistically indistinguishable from kNN-VC and SoundStorm in both naturalness and speaker similarity. confidence: high relevance: high - claim_id: linearvc_orthogonal_content_speaker_subspaces role: supports claim: Content and speaker identity information reside in orthogonal low-dimensional subspaces within the same SSL layer, enabling voice style transfer through geometric manipulation. source: §4, Table 2; §5.2, Figure 4 evidence: Constraining the linear transformation to rotation and reflection only achieves EER 27.7% vs. 31.8% for unconstrained; adding translation alone gives only EER 7.7% with intelligibility maintained (CER 2.9%); SVD factorization at rank 16 achieves CER < 4% while speaker similarity requires rank ~100. confidence: high relevance: high - claim_id: linearvc_naturalness_similarity_tradeoff role: complicates claim: High naturalness in voice conversion does not imply high speaker similarity; these objectives can trade off sharply depending on the system design. source: §3.2, Table 1 evidence: FreeVC achieves the highest naturalness (71.1 MUSHRA) among all systems but the lowest speaker similarity (EER 10.5% vs. 33.6% for LinearVC and 38.9% for kNN-VC), indicating perceptual smoothness and target-speaker fidelity are partly in tension. confidence: high relevance: high - claim_id: linearvc_eer_coarse_similarity_proxy role: refines claim: EER as an objective speaker similarity metric provides only coarse correspondence with perceived speaker similarity in voice conversion evaluation. source: §3.2 evidence: The paper notes small EER differences do not reliably track subjective similarity ratings; LinearVC and kNN-VC have comparable subjective similarity (67.5 vs. 67.2) but different EERs (33.6 vs. 38.9). confidence: medium relevance: high limitations: - The analysis is restricted to WavLM-Large (layer 6) on English LibriSpeech; whether the orthogonal subspace structure holds for other SSL models, layers, or languages is left untested, and the shared pretrained HiFi-GAN vocoder makes it difficult to isolate conversion quality from synthesis quality. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-0464 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - VC - codec architecture: - autoregressive-LM relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - autoregressive_generation claims: - claim_id: explicit_mutual_information_minimisation_at_the_codec_embedding role: supports claim: Explicit mutual information minimisation at the codec-embedding level provides effective prosody-timbre disentanglement for voice conversion. source: §3.5, Table 3 evidence: Removing the MI loss (L_MI) from the full system leads to a notably higher normalised F0 distance in the prosody-from-source scenario (3.28 vs. 2.82), while quality and timbre metrics change only modestly, isolating prosody control as the primary benefit of the MI objective. confidence: high relevance: low - claim_id: in_context_learning_codec_lms_can_serve_as role: supports claim: In-context learning codec LMs can serve as controllable VC backbones when augmented with prosody-disentangling encoder modules. source: §3.4, §3.5, Table 2, Table 3 evidence: The proposed system builds on VALL-E X's ICL mechanism and outperforms VALL-E X in speaker similarity (ASV 0.91 vs. 0.84), intelligibility (WER 0.101 vs. 0.115), naturalness (MOS 4.36 vs. 4.19), and prosody alignment (F0 distance 2.70 vs. 3.10) in the prompt-based scenario. confidence: high relevance: high - claim_id: prosody_disentanglement_at_the_codec_level_introduces_a role: complicates claim: Prosody disentanglement at the codec level introduces a small trade-off in absolute codec reconstruction fidelity compared to the unmodified encoder. source: §3.3, Table 1 evidence: PACE's ASV score (0.662) and NISQA score (3.98) are lower than the baseline EnCodec encoder (0.681, 4.17), though the gap does not substantially affect system-level VC performance. confidence: high relevance: low - claim_id: prosody_from_source_and_prosody_from_prompt_are role: refines claim: Prosody-from-source and prosody-from-prompt are distinct capability axes in voice conversion; systems strong at one do not automatically handle the other. source: §3.5, Table 3 evidence: VALL-E X supports only prosody-from-prompt and is excluded from the source-prosody evaluation; TriAAN-VC and ProsoVC support only source-prosody and are excluded from the prompt-prosody evaluation. Only the proposed system is evaluated in both modes. confidence: high relevance: low limitations: - All evaluation is conducted on LibriTTS-clean-100 and test-clean, a relatively clean single-domain corpus with 247 speakers. Generalisation to noisy environments, expressive or emotional speech, or cross-lingual settings is not tested. - The 54-hour training dataset is modest for a codec language model approach; it is unclear whether the disentanglement quality degrades with longer or more expressive source utterances. No code or demo is reported, limiting reproducibility. The paper does not ablate the number of RVQ codebooks or the sensitivity of the MI-minimisation trade-off weight (lambda_MI), leaving the robustness of the disentanglement objective undercharacterised. Prosody is operationalised solely through f0 and UV binary flags; richer prosodic dimensions such as energy, speaking rate, and phrase-level structure are not captured. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-0469 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS - evaluation architecture: - transformer-enc-dec relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - transformer_encoder_decoder claims: - claim_id: phoneme_based_tts_outperforms_character_based_multilingual role: supports claim: Phoneme-based TTS training substantially outperforms character-based multilingual approaches in intelligibility for low-resource South Asian languages, as measured by WER, CER, and subjective MOS. source: §4.2, §4.3 evidence: Tacotron 2 + WaveGlow achieves WER 16.1% versus MMS's 31.4% on Punjabi male, and Tacotron 1 achieves WER 13.1% versus MMS's 25.6% on Urdu male. confidence: high relevance: high - claim_id: transliteration_workarounds_degrade_tts_quality role: complicates claim: Script-mismatch workarounds such as transliteration introduce cumulative errors that degrade TTS quality below what a dedicated script-native model achieves. source: §1, §5 evidence: Meta MMS lacks native Punjabi Shahmukhi support and required an additional Shahmukhi-to-Gurumukhi transliteration step via ChatGPT before synthesis, contributing to its substantially higher WER and CER relative to the dedicated Shahmukhi-native Tacotron models. confidence: medium relevance: low - claim_id: multilingual_tts_fails_on_numerals_and_symbols role: supports claim: Large-scale multilingual TTS models fail to synthesize numerals and language-specific symbols correctly in languages whose text processing was not explicitly handled at training time, causing systematic intelligibility failures. source: §4.2, §4.3 evidence: MMS failed completely to synthesize numbers and symbols in both Punjabi and Urdu, severely degrading intelligibility relative to the dedicated Tacotron systems with script-specific text-analysis modules. confidence: high relevance: low - claim_id: phoneme_level_metrics_expose_weaknesses_mos_misses role: refines claim: Phoneme-level evaluation metrics (DRT/MRT) expose word-final articulatory weaknesses in TTS systems that aggregate perceptual scores like MOS do not adequately capture. source: §4.3.1 evidence: DRT/MRT phoneme-feature accuracy testing showed MMS weakest at word-final positions across voicing, nasality, aspiration, and sibilation in Punjabi, distinctions not visible in the aggregate MOS or SUS scores. confidence: medium relevance: high - claim_id: simpler_architecture_matches_larger_for_lowresource_transcription role: supports claim: For low-resource encoder-decoder TTS, simpler architectures can match or exceed larger ones in transcription accuracy while offering substantially lower inference latency. source: §4.1, §4.2 evidence: Tacotron 1 (6.9M params, RTF 0.17) achieved the best Urdu male transcription accuracy (WER 13.1%, CER 4.0%), while the larger Tacotron 2 + WaveGlow (296M params combined) had RTF 3.54. confidence: medium relevance: low limitations: - The Punjabi corpus covers only the Majhi accent, and the MMS comparison is limited since MMS training data details are unavailable, preventing MCD-DTW and SpeechBERTScore computation for MMS. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-0554 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS architecture: - flow-matching - GAN relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - gan_adversarial_generation - flow_matching_generation claims: - claim_id: rapflow_straight_trajectory_better_for_consistency role: supports claim: Enforcing velocity-field self-consistency on straight flow-matching trajectories yields more effective few-step TTS than applying consistency constraints to curved diffusion-model trajectories. source: §1, §3.1, Table 1 evidence: RapFlow-TTS builds consistency constraints on Matcha-TTS's straight flow-matching trajectories and at 2 NFE (MOS 3.89) exceeds Comospeech's diffusion-based consistency distillation at 2 NFE (MOS 3.19) on LJSpeech. confidence: high relevance: high - claim_id: rapflow_adversarial_learning_most_impactful role: supports claim: Adversarial learning at the mel-spectrogram level is the single most impactful technique for improving few-step consistency TTS naturalness, outperforming all other regularisation methods individually. source: §3.2, §4.3, Table 2 evidence: Ablation confirms adversarial learning yields the largest NISQA improvement (3.78 to 4.19 on LJSpeech) among all improved techniques tested (encoder freeze, shared dropout, Huber loss, delta scheduling). confidence: high relevance: high - claim_id: rapflow_consistency_models_dont_benefit_from_more_steps role: supports claim: Consistency models for TTS do not benefit from increasing the number of inference steps beyond what they were trained for, and can degrade with more steps. source: §4.3, Table 2 evidence: The paper documents that more NFE steps do not consistently improve quality for RapFlow-TTS's consistency-trained model, a characteristic behaviour distinct from standard flow-matching models. confidence: high relevance: low - claim_id: rapflow_two_stage_recipe_enables_2step_realtime role: supports claim: Combining a straight flow pre-training stage with consistency fine-tuning and multi-segment objectives enables 2-step TTS at real-time factors comparable to non-autoregressive models such as FastSpeech2. source: §3.1, §4.2, Table 1 evidence: RapFlow-TTS+ achieves RTF 0.031 at 2 NFE, matching FastSpeech2's speed, while reaching MOS 4.01 versus FastSpeech2's 3.35 on LJSpeech. confidence: high relevance: high limitations: - Quality at 2 NFE remains slightly below ground truth (MOS 4.01 vs GT 4.42 on LJSpeech). - The system is not zero-shot; it conditions on speaker identity via training rather than a reference prompt, and extension to large-scale data and zero-shot capability is left to future work. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-0648 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: automated_annotation_exceeds_human_consistency role: supports claim: Automated multimodal annotation pipelines can achieve annotation consistency significantly exceeding that of human raters on emotional speech datasets, at substantially lower cost per hour. source: §2.3, Table 1 evidence: MIKU-PAL achieves Fleiss kappa of 0.93-0.95 on MELD and IEMOCAP versus human annotator kappa of only 0.40-0.43, at roughly 1/20th the cost (~$0.50/hour vs. ~$10/hour). confidence: high relevance: high - claim_id: finegrained_consistent_datasets_improve_emotional_tts role: supports claim: Fine-grained emotion datasets with a larger category space and higher annotation consistency lead to measurable improvements in downstream emotional TTS quality compared to smaller manually annotated alternatives. source: §3.2, Table 3 evidence: Fish-Speech fine-tuned on MIKU-EmoBench (131.2h, 26 emotion categories) achieves MOS 4.12 and emotion similarity 0.92 on the MELD test set, versus 3.91 MOS and 0.88 emotion similarity for the baseline. confidence: medium relevance: high - claim_id: visual_preprocessing_boosts_multimodal_emotion_accuracy role: supports claim: Visual preprocessing is a substantial contributor to multimodal emotion recognition accuracy, as removing it causes a measurable drop in classification performance. source: §2.1 evidence: Ablating active-speaker visual preprocessing (via TalkNet-based speaker identification) reduces MELD classification accuracy by 25.6% relative to the full pipeline. confidence: medium relevance: low - claim_id: limited_emotion_taxonomies_cause_benchmark_confusion role: complicates claim: The limited emotion category sets in standard SER benchmarks introduce systematic classification confusion for emotionally adjacent categories that are not psychologically distinct. source: §2.3 evidence: The paper argues standard benchmarks using Ekman's six basic emotions (IEMOCAP, MELD) cannot capture nuanced transitions confirmed by t-SNE visualization on MIKU-PAL's 26-category outputs (e.g., admiration to love, joy to satisfied to excitement), motivating the psychologically grounded 26-category taxonomy. confidence: medium relevance: low limitations: - MIKU-PAL accuracy (58.6-68.5%) remains below human accuracy on IEMOCAP (72.9%), and current SER models cannot reliably classify all 26 fine-grained emotion categories, limiting full evaluation of MIKU-EmoBench. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-0656 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - VC architecture: - hybrid relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - hybrid_generation claims: - claim_id: eegvc_crossmodal_alignment_enables_vc role: supports claim: Cross-modal feature alignment between neural signals and speaker embeddings can enable voice conversion without any target-speaker voice data. source: §3.1.2, §4.4 evidence: The EEG-voice feature alignment module, trained with embedding MSE and speaker classification losses, produces speaker embeddings from EEG that drive FreeVC-based conversion to zero-shot quality (Naturalness MOS 4.00, Consistency obj 0.8026 for unseen speakers). confidence: medium relevance: high - claim_id: eegvc_nonspeech_biometric_encodes_identity role: supports claim: Non-speech biometric signals can encode speaker-identity information sufficient to guide voice timbre conversion. source: §4.4.1, Figure 2 evidence: t-SNE visualisation shows synthesised speech clusters align with reference audio per speaker, and Homogeneity scores (0.9437-0.9465) exceed the FreeVC speech-only baseline (0.9371), indicating EEG features encode timbre-discriminative information. confidence: medium relevance: low - claim_id: eegvc_requires_speech_pretraining_stage role: complicates claim: Zero-shot voice conversion from EEG signals requires a large-scale speech-only pre-training stage to compensate for the scarcity and noise of paired EEG-speech data. source: §3.2 evidence: The three-stage curriculum first pre-trains on VCTK (Stage I, speech only), then aligns EEG to pre-trained speaker embeddings (Stage II), before joint fine-tuning (Stage III); the authors explicitly attribute feasibility to leveraging the pre-trained VC model's representations. confidence: medium relevance: low - claim_id: eegvc_evaluation_limited_by_paired_data role: complicates claim: Evaluation of EEG-driven voice conversion is fundamentally limited by the availability of paired EEG-speech corpora at the scale needed for generalisation. source: §4.1, §5 evidence: The entire EEG evaluation uses the Single-Word-Production Dutch-iBIDS dataset (10 speakers, single-word utterances); seen-speaker training uses 80% of this data, and the authors note that more data will improve performance. confidence: medium relevance: low limitations: - The evaluation dataset contains only 10 speakers producing single words in Dutch; the baseline comparison (FreeVC) is not equivalent since it uses a target-speaker voice prompt rather than EEG, and no intelligibility metrics (WER/CER) are reported. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-0706 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - SCA architecture: [] relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: [] claims: - claim_id: llm_generated_qa_correlates_with_human_authored_qa role: supports claim: Automated LLM-based QA generation can produce paralinguistic evaluation sets that correlate closely with human-authored ones for assessing speech LLMs. source: §4, Table 2 evidence: Qwen2-Audio-7B-Instruct achieves 60.28 (LLM-generated QA) versus 59.46 (human-generated QA) under the ChatGPT judge with Prompt 2, and 56.82 versus 54.33 under the Llama-70B judge, both differences small and directionally consistent. confidence: medium relevance: high - claim_id: dual_ser_consistency_filtering_reduces_annotation_noise role: supports claim: Combining categorical and dimensional emotion recognition with consistency filtering reduces annotation noise in in-the-wild speech data condensation. source: §3.1, Figure 3 evidence: The dual SER consistency condition (categorical sentiment class aligned with valence score) applied on SG TV/Movie data achieved the highest UWA of 33.65% at valence thresholds x=0.5, y=0.4, outperforming single-paradigm labelling and improving class balance. confidence: medium relevance: high - claim_id: speech_llms_weak_at_empathetic_reasoning role: complicates claim: Current speech LLMs exhibit weak performance on contextual empathetic reasoning even when evaluated against well-formed paralinguistic QA. source: §4, §5 evidence: Evaluation of Qwen2-Audio-7B-Instruct on the CPQA set reveals limitations in handling empathetic reasoning tasks, identified as a motivating gap requiring both better data and more robust models. confidence: medium relevance: low - claim_id: automatic_qa_generation_introduces_biases_needing_filtering role: complicates claim: Scalable automatic QA generation from speech introduces systematic biases that require post-filtering to produce usable evaluation data. source: §3.2 evidence: ChatGPT-generated QA contained repetitive question variants (multiple paraphrases asking about reasons behind emotion) and irrelevant questions assuming a text transcript was available, requiring keyword-based post-filtering to remove them. confidence: medium relevance: low limitations: - The evaluation benchmark is small (480 samples, 6.5 hours) drawn from a single domain (Singaporean YouTube channels), and the entire framework is validated against a single speech LLM (Qwen2-Audio-7B-Instruct). caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-0739 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - SCA - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: internal_interrupt_detection_beats_vad_based_fdsds role: supports claim: Full-duplex spoken dialogue systems that integrate interruption detection internally rather than relying on external VAD modules show systematically higher interrupt success rates under challenging conditions. source: §3.2, Table 4 evidence: Moshi, which uses a built-in full-duplex mechanism, achieves SRRate under Easy interruptions of 61.7% versus 11.3% for Freeze-Omni and 26.9% for VITA-1.5, both of which rely on external VAD. confidence: high relevance: high - claim_id: latency_and_interruption_robustness_tradeoff role: complicates claim: 'Real-time responsiveness (first-speech-emit delay) and interruption robustness trade off against each other: systems optimised for low latency may not handle mid-utterance interruptions reliably.' source: §3.2, Table 4 evidence: Freeze-Omni achieves the lowest first-speech-emit delay (449-515 ms versus 2,020-4,155 ms for Moshi) yet handles interruptions far less reliably than Moshi. confidence: medium relevance: high - claim_id: background_noise_worse_than_gap_noise_for_fdsds role: supports claim: Background noise degrades response quality in full-duplex spoken dialogue systems more severely than discrete noise events inserted into silence gaps. source: §3.2, Table 5 evidence: Continuous background noise (Noise-bg) at low SNR degrades subjective response-quality scores across all three tested systems, while gap noise inserted into silence periods has comparatively little effect, with Freeze-Omni's high VAD threshold blocking most spurious gap events. confidence: medium relevance: high - claim_id: automated_scoring_and_perplexity_are_complementary_not_sufficient role: refines claim: Automated LLM-based scoring and conditioned perplexity provide complementary signals for evaluating spoken dialogue quality, but neither alone captures interruption-handling behaviour. source: §2.3, §3.2 evidence: FD-Bench combines GPT-4o-based six-dimension subjective scoring with Llama-3.3-70B conditioned perplexity (c-PPL) alongside dedicated timing metrics (IRD, FSED) because response-quality scoring alone does not capture interruption success or timing behavior. confidence: medium relevance: high - claim_id: fdsds_far_from_human_robustness_under_interruption role: supports claim: Current full-duplex spoken dialogue systems remain far from human-level robustness under frequent interruptions, suggesting the capability remains an open engineering and research challenge. source: §3.2, §5 evidence: All three tested open-source FDSDS (Moshi, Freeze-Omni, VITA-1.5) show substantial performance degradation under frequent interruptions and noisy conditions, with no system approaching reliable interrupt handling across all tested scenarios. confidence: medium relevance: high limitations: - All reported results use speech synthesised by CosyVoice2 only; the ChatTTS and F5-TTS subsets are present in the corpus but results are not reported in the paper, and evaluation is English-only. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-0762 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS architecture: - flow-matching relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: - flow_matching_generation claims: - claim_id: swedish_psychometric_accentedness_calibration role: supports claim: Recursive psychometric function estimation can establish a perceptually calibrated mapping from a continuous TTS conditioning parameter to perceived accentedness, yielding multiple distinguishable perceptual levels. source: §3.2, §4.2 evidence: The sigmoid psychometric function for English-accented Swedish achieves R² = 0.983, and MAP-SV testing identified at least 9 perceptually distinguishable EA levels within [0,1], versus only 3 for MAP-EN. confidence: high relevance: high - claim_id: swedish_transcription_accentedness_alignment role: supports claim: Misalignment between transcription style and accentedness level is consistently dispreferred by listeners, regardless of which misaligned combination is presented. source: §3.3, §4.3, Table 1 evidence: TEA-ALIGN listening tests with 16 Swedish native speakers show aligned transcription+PEA conditions receive significantly more best votes than misaligned ones (χ²(2, N=630) = 50.2, p < .05). confidence: high relevance: high - claim_id: swedish_insertion_type_dependent_preference role: supports claim: Listener preference for accentedness in intrasentential English insertions is insertion-type-dependent rather than uniform, requiring fine-grained control beyond binary language switching. source: §3.4, §4.4, Table 1 evidence: 'PREF-VAR results with 32 listeners show no single PEA setting dominates across insertion types: well-known names are preferred in Swedish while English technical terms are preferred in English.' confidence: medium relevance: high - claim_id: swedish_asymmetric_accent_sensitivity role: supports claim: 'Perceptual sensitivity to accent direction is asymmetric: listeners can distinguish more granular levels of English-accentedness in Swedish speech than of Swedish-accentedness in English speech.' source: §4.2 evidence: MAP-SV identified at least 9 distinguishable EA levels for English-accented Swedish, while MAP-EN found only 3 for Swedish-accented English. confidence: medium relevance: high limitations: - Single speaker and single Swedish/English language pair; the underlying flow-matching backbone (Matcha-TTS) is used unmodified, so generalisation to other flow-matching architectures or multi-speaker settings is untested. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-0779 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: asr_metrics_uncorrelated_with_human_math_intelligibility role: complicates claim: Standard ASR-based metrics (CER, BLEU, ROUGE) correlate near-zero with human transcription accuracy for TTS of mathematical expressions, making them insufficient proxies for intelligibility in this domain. source: §5.3 evidence: TTS-ASR cascade metrics (CER=0.10, BLEU=0.66, ROUGE=0.84, averaged across models) correlate at less than 0.15 absolute with human intelligibility metrics (Count-of-Correct, LCER). confidence: high relevance: high - claim_id: math_tts_intelligibility_varies_strongly_by_category role: supports claim: TTS intelligibility for mathematical content is strongly dependent on expression category, with numeric expressions near ceiling and summation or matrix expressions substantially degraded, even for state-of-the-art commercial systems. source: §4, Table 2 evidence: Numerics achieve MOS 4.7-4.9 and Count-of-Correct approximately 97-100% across all five commercial TTS systems, while Summation scores MOS 2.3-3.5 with Count-of-Correct only 40-59%, and Matrices score MOS 3.1-3.6 with Count-of-Correct only 17-29%. confidence: high relevance: high - claim_id: no_single_tts_model_dominates_across_math_categories role: refines claim: No single commercial TTS model consistently outperforms others across all mathematical expression categories, indicating that model selection for math-aware TTS must be category-specific. source: §4, Table 2 evidence: Across the eight evaluated mathematical expression categories, the best-performing commercial system (among Amazon Polly, Azure, Google, OpenAI, Seamless) varies by category, with no model dominating on all eight. confidence: medium relevance: low - claim_id: human_expert_math_rendition_near_perfect_vs_tts_gap role: supports claim: Human expert rendition of mathematical expressions achieves near-perfect intelligibility (97% CoC), establishing a large gap relative to all evaluated TTS systems for most expression categories. source: §4, Table 3 evidence: In the L2 MUSHRA-inspired comparison against a hidden human-expert reference, TTS models show difference scores of 24-73 points on a [0,100] scale for all categories except Numerics and Fractions, with human expert Count-of-Correct at 97% versus substantially lower TTS averages. confidence: high relevance: high - claim_id: llm_pronunciation_correctness_not_the_bottleneck role: refines claim: LLM-generated pronunciation text does not significantly affect TTS intelligibility when the pronunciation is correct, indicating that the bottleneck lies in the TTS prosody rather than the text normalization step. source: §4 evidence: ANOVA on LaTeX Character Error Rate shows that LLM choice (QWEN2.5-7B vs. GPT-4) for generating pronunciation text is not a statistically significant factor once pronunciation is verified correct, while TTS model and expression category are both significant (p < 0.05). confidence: medium relevance: low limitations: - Only English mathematical expressions are evaluated, and fine-tuning TTS systems specifically for mathematical content is left as future work. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-0854 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS architecture: - GAN - flow-matching - transformer-enc-dec relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - transformer_encoder_decoder - gan_adversarial_generation - flow_matching_generation claims: - claim_id: postprocessor_effectiveness_degrades_in_low_resource_settings role: supports claim: Generative postprocessing models for spectrogram enhancement lose effectiveness in low-resource settings because the acoustic model produces stronger artifacts that are underrepresented in the available training pairs. source: §1, §2.1 evidence: Low-resource speakers (under 15 minutes of audio) produce stronger acoustic-model artifacts while yielding too few (synthetic, real) training pairs for the postprocessor to learn to compensate. confidence: medium relevance: low - claim_id: subsampled_high_resource_data_simulates_low_resource_artifacts role: supports claim: Simulating low-resource speaker conditions from subsampled high-resource data provides a viable source of artifact-realistic training pairs for spectrogram postprocessors, yielding consistent naturalness improvements across both GAN and flow-matching architectures. source: §2.2, §4, Table 2 evidence: GAN with the proposed data generation reaches MUSHRA 74.8 vs. 72.3 standard (p ≤ 0.01), and CFM reaches 79.8 vs. 79.0 standard, both improving over the ForwardTacotron baseline of 50.9. confidence: high relevance: high - claim_id: cfm_postprocessor_higher_naturalness_smaller_relative_gain role: refines claim: Flow-matching postprocessors achieve higher absolute naturalness than GAN-based postprocessors on low-resource speakers, but show smaller and statistically insignificant gains from the proposed data generation strategy compared to GAN-based postprocessors. source: §4, Table 2 evidence: CFM-proposed (79.8) is not statistically significantly better than CFM-standard (79.0), while the GAN-proposed vs. GAN-standard gap (74.8 vs. 72.3) is significant at p ≤ 0.01. confidence: high relevance: high - claim_id: noise_augmentation_does_not_improve_spectrogram_naturalness_with_gt_prosody role: complicates claim: Noise augmentation of the acoustic model does not improve spectrogram naturalness when ground-truth prosody is used at evaluation, as its primary benefit is prosody prediction rather than spectrogram texture quality. source: §4, Table 2 evidence: The noise-augmentation baseline (MUSHRA 45.9) underperforms the plain ForwardTacotron baseline (50.9) in this ground-truth-prosody evaluation setup. confidence: medium relevance: high - claim_id: mushra_insufficient_to_detect_synthetic_exceeding_reference_quality role: complicates claim: Multi-stimulus hidden-reference tests alone may be insufficient to detect when a TTS system exceeds vocoded ground-truth quality; ranking-by-elimination tests can reveal preferences for synthetic over reference conditions that MUSHRA-style scoring cannot capture. source: §3.3, §4 evidence: The CFM-proposed system was preferred over the reference in 26.6% of RBE (Plackett-Luce) rankings, a signal not directly visible in the standard MUSHRA comparison. confidence: medium relevance: high limitations: - Evaluation uses ground-truth prosody at inference, so it is unclear how the approach performs in full end-to-end synthesis with predicted prosody; only two low-resource speakers (one male, one female) are tested. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-0902 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - VC architecture: - VAE - GAN relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: - gan_adversarial_generation - variational_latent_generation claims: - claim_id: explicit_voice_quality_conditioning_enables_controllable_phonation role: supports claim: Explicit conditioning on acoustic voice quality features (CPPS, H1-H2, H1-A3, creakiness) enables controllable manipulation of phonation type in a VC system. source: §3.2, §4.1 evidence: The system demonstrates controllable modification of all five conditioned features across a range of -3 to 3 standard deviations from the mean. confidence: high relevance: low - claim_id: breathy_voice_perceived_more_intimate_and_invested role: supports claim: Breathy voice is perceived as more intimate and more invested than modal voice when paralinguistic content is held constant across speakers. source: §5.2, Table 1 evidence: In the 25-participant listening test, breathy voice was rated significantly more intimate (p=.01) and more invested (p<.01) than modal voice. confidence: high relevance: high - claim_id: creaky_voice_perceived_less_intimate_and_positive role: supports claim: Creaky voice is perceived as less intimate and less positive than modal voice, consistent with prior findings using natural speech stimuli. source: §5.2, Table 1 evidence: Creaky voice was rated significantly less intimate (p=.04) and less positive (p<.01) than modal voice in the same listening test. confidence: high relevance: high - claim_id: lightweight_feature_encoders_sufficient_for_voice_quality_control role: supports claim: Fine-tuning an existing VC backbone with lightweight feature encoders is sufficient to achieve intuitive control over multiple acoustic correlates of voice quality without degrading the base model's conversion capability. source: §3.3, §4.1 evidence: Five affine voice-quality encoders add only about 12K parameters over the base FreeVC model (39.4M total) while enabling controllable phonation manipulation. confidence: medium relevance: low limitations: - No MUSHRA or absolute MOS for speech naturalness is reported; evaluation is focused on paralinguistic perception rather than synthesis fidelity. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-0973 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: english_trained_mos_predictors_underperform_on_spanish role: supports claim: Automatic MOS predictors trained primarily on English data underperform on Spanish TTS, and language-specific fine-tuning provides meaningful improvement even with small datasets. source: §4.2, Table 2 evidence: Out-of-box NISQA v1.0 (English-trained) reaches PCC=0.71/MAE=0.99 on Spanish, improving to PCC=0.73/MAE=0.81 after fine-tuning on 3,139 Spanish training samples. confidence: high relevance: high - claim_id: low_level_ssl_features_more_informative_for_naturalness_prediction role: supports claim: Low-level local acoustic features from SSL encoder layers are more informative for naturalness prediction than higher-level contextual representations from deeper transformer blocks. source: §4.2, Figure 2 evidence: In DenseMOS's learned layer-averaging weights, the wav2vec 2.0 CNN encoder output (25 ms local resolution) receives the highest weight, outweighing the deeper transformer blocks. confidence: medium relevance: high - claim_id: mos_label_distribution_skew_biases_predictor_toward_mean role: complicates claim: The scarcity of samples rated near MOS 4.0 in evaluation datasets introduces a systematic prediction bias toward the dataset mean, suggesting label distribution matters as much as dataset size for MOS predictor quality. source: §4.2 evidence: All automatic MOS models in this study struggle to predict scores near 4.0, which is underrepresented in the 4,326-sample Spanish dataset, producing a systematic bias toward the mean. confidence: high relevance: high - claim_id: lightweight_ssl_downstream_models_match_finetuned_specialists role: supports claim: Lightweight downstream models trained on frozen SSL representations can achieve MOS prediction performance comparable to fine-tuned specialist models, despite having fewer than half the parameters. source: §4.2, Table 2 evidence: DenseMOS (115K params, frozen wav2vec 2.0) reaches PCC=0.62/MAE=0.81, comparable to fine-tuned NISQA (305K params) at PCC=0.73/MAE=0.81. confidence: medium relevance: high limitations: - Most audio samples received only a single rating, making inter-rater reliability estimates noisy at the individual-item level; the dataset is geographically skewed toward Argentine Spanish and does not include modern zero-shot or LLM-based TTS systems. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-0984 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - codec - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: neural_codecs_below_1100bps_degrade_intelligibility role: supports claim: Neural speech codecs operating below approximately 1,100 bps degrade in intelligibility relative to uncoded speech, even when subjective quality metrics suggest acceptable performance. source: §4.1 evidence: All neural codecs tested below roughly 1,100 bps show significant DRT accuracy degradation relative to the uncoded reference; SemantiCodec at 340 bps achieves only 76-78% DRT accuracy, the lowest among neural codecs tested. confidence: high relevance: high - claim_id: wer_not_valid_intelligibility_proxy_for_codec_benchmarking role: contradicts claim: WER derived from large-vocabulary ASR systems is not a valid proxy for subjective speech intelligibility in codec benchmarking on isolated word stimuli. source: §4.3 evidence: WER (via Whisper) shows no meaningful correlation with subjective DRT scores (r = -0.11 averaged, -0.15 disaggregated), and Whisper itself yields 19-25% WER on the clean original reference signal. confidence: high relevance: high - claim_id: stoi_estoi_correlate_aggregate_but_miss_finegrained_variation role: refines claim: STOI and ESTOI correlate well with subjective DRT intelligibility scores when averaged across talker gender and wordlists, but fail to capture the finer-grained variation that subjective tests reveal. source: §4.3 evidence: 'STOI achieves Pearson r=0.958 and ESTOI r=0.890 averaged over gender and wordlist, but correlation drops substantially at the disaggregated level (STOI: 0.595, ESTOI: 0.499).' confidence: high relevance: high - claim_id: talker_gender_affects_intelligibility_independent_of_codec role: complicates claim: Talker gender significantly affects subjective intelligibility scores independently of codec condition, and this variation is not reliably reflected in objective intelligibility metrics. source: §4.1, §4.2 evidence: The female reference signal scores lower than the male (M=85 vs. M=91.87) and this gender gap propagates through codec conditions, statistically significant in a Linear Mixed-Effects Model (p < .0001). confidence: medium relevance: high limitations: - The female reference signal shows consistently lower DRT scores than the male reference, suggesting potential issues with the audio materials rather than codec effects, which limits confidence in codec-level conclusions for female talkers; the benchmark is restricted to English and 16 kHz stimuli. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1066 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS architecture: - diffusion relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - diffusion_generation claims: - claim_id: ebm_score_training_singlestep role: supports claim: Score-based training objectives for energy-based models can reduce TTS inference to a single step, bypassing the many-iteration MCMC requirement of NCE-trained EBMs. source: §4.2, Table 2, Table 3 evidence: Both SSM- and delta-loss-trained EBMs obtain their best UTMOSv2 and MOS results after just 1 inference step, versus the 100-step MCMC schedule needed by NCE-trained EBMs from prior work. confidence: high relevance: high - claim_id: ebm_training_objective_shapes_score_geometry role: supports claim: The choice of training objective for score-based generative models materially affects the geometry of learned score functions and their suitability for first-order iterative inference. source: §3.2 evidence: Delta loss, which explicitly enforces linear score paths between hypothesis and reference, degrades sharply at 100 steps (UTMOSv2 drops from 3.703 to 2.937) while SSM, which does not enforce this geometry, remains comparatively stable. confidence: medium relevance: high - claim_id: ssm_fm_formal_equivalence role: supports claim: Sliced score matching and flow matching objectives are formally equivalent at a single interpolation timestep, suggesting a deeper structural connection between energy-based and flow-based TTS training regimes. source: §3.3 evidence: The paper derives that its delta-loss objective (equation 9) is formally equivalent to the flow-matching objective (equation 10) when t=0, connecting EBM score training to the continuously-interpolated vector field used in rectified-flow-style TTS. confidence: high relevance: low - claim_id: ebm_mos_utmos_divergence role: complicates claim: Automatic MOS predictors and subjective listening scores can diverge for EBM-based TTS systems, limiting the reliability of UTMOSv2 as a sole quality proxy in non-standard synthesis paradigms. source: §4.2, §4.3 evidence: Delta loss produces a low UTMOSv2 score (2.959) but a competitive subjective MOS (3.95 ± 0.05), with far fewer low listener ratings than SSM despite comparable mean MOS. confidence: medium relevance: high limitations: - Experiments use only single-speaker LJSpeech with a Tacotron 2 + mel + HiFi-GAN backbone substantially behind current codec-based or DiT-based practice, so it is untested whether the flow-matching-equivalence finding transfers to modern flow-matching TTS architectures. - The subjective evaluation uses only 3 listeners and 100 test clips, giving insufficient statistical power to draw strong conclusions. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1081 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - VC architecture: - GAN relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - gan_adversarial_generation claims: - claim_id: sncr_content_restoration_compensates_in_degradation role: supports claim: Explicit content-restoration mechanisms can compensate for linguistic information degraded by instance normalization in speaker disentanglement, preserving intelligibility without SSL feature extractors. source: §2.1, Table 2 evidence: The SNCR dual restoration loss, which decomposes the instance-normalization residual into content-relevant and content-irrelevant parts via phoneme-entropy constraints, reduces WER from 5.57% (without SNCR) to 3.78% compared to a baseline trained with neither IN nor dual restoration loss. confidence: high relevance: low - claim_id: sncr_gan_matches_diffusion_lower_cost role: supports claim: GAN-based zero-shot voice conversion systems can match the speaker similarity of diffusion-based systems at substantially lower parameter count and faster inference. source: §3.2, Table 1 evidence: The proposed 50M-parameter GAN system achieves SECS 0.25 versus DDDM-VC's SECS 0.27, while running at RTF 0.11 versus 0.29 and using 366M fewer parameters; naturalness (nMOS 3.86 vs. 3.52) also favors the GAN system. confidence: high relevance: high - claim_id: sncr_attention_discriminator_improves_quality role: supports claim: Integrating self-attention layers into CNN-based discriminators improves adversarial training quality for mel-spectrogram reconstruction in voice conversion. source: §3.3, Table 2 evidence: Ablation with the discriminator's self-attention layers removed (replaced with a standard CNN discriminator) degrades nMOS and increases WER compared to the full system. confidence: medium relevance: high - claim_id: sncr_speaker_similarity_inflated_by_overlap role: complicates claim: Speaker similarity metrics in zero-shot VC evaluations can be inflated when the test set speakers overlap with a baseline system's training set, making fair cross-system comparison difficult. source: §3.2 evidence: VQMIVC achieves SECS 0.23 (close to the proposed system's 0.25) but was trained on VCTK, which overlaps with the VCTK test speakers; its WER of 21.4% reveals conversions are not actually intelligible despite the seemingly competitive SECS. confidence: high relevance: high limitations: - The system trains and evaluates on English only (LibriLight and VCTK); the ablation evaluates components one at a time, leaving interactions between SNCR and consistency loss unexplored. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1084 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS architecture: - autoregressive-LM relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - autoregressive_generation claims: - claim_id: mamba_based_sequence_models_can_match_or_exceed role: supports claim: Mamba-based sequence models can match or exceed the quality of larger Transformer-based TTS systems while enabling real-time streaming inference on CPU hardware. source: §4.5, §4.6, Table 1 evidence: SMAM+MLM (26M params) achieves MOS 4.02 and CER 2.73%, matching Lee et al. (2024) at 263M params (MOS 4.00, CER 4.01%) while reducing first-token latency from 26.5s to 0.065s on a single-threaded CPU. confidence: high relevance: high - claim_id: iterative_depthwise_refinement_of_rvq_tokens_substantially_improves role: supports claim: Iterative depthwise refinement of RVQ tokens substantially improves codec TTS quality over single-pass parallel depth prediction. source: §4.7, Table 1 evidence: Replacing MLM depthwise decoding with a single-pass no-masking baseline (SMAM+noMLM) causes a significant drop in all quality metrics (MOS from 4.02 to 3.89, CER from 2.73% to 4.12%, UTMOS from 4.13 to 3.83) with negligible change in RTF and latency. confidence: high relevance: high - claim_id: objective_speaker_similarity_metrics_based_on_embedding_cosine role: complicates claim: Objective speaker similarity metrics based on embedding cosine distance do not reliably predict subjective speaker similarity as judged by human listeners. source: §4.6, Table 1 evidence: SMAM+MLM scores SECS 0.816 (below Lee et al.'s 0.863) but achieves higher SMOS of 3.36 vs. 3.27, indicating a divergence between embedding-space distance and perceptual similarity that has practical implications for zero-shot TTS evaluation. confidence: high relevance: high - claim_id: depthwise_decoding_strategies_for_rvq_present_an_explicit role: supports claim: Depthwise decoding strategies for RVQ present an explicit quality-speed trade-off that system designers can exploit based on deployment constraints. source: §3.3, §4.5, §4.6, Table 1 evidence: SMAM+MLM (iterative, 3 passes) achieves MOS 4.02 and RTF 0.701, while SMAM+INR (single forward pass) achieves MOS 3.97 and RTF 0.568, demonstrating a consistent quality-speed trade-off across both objective and subjective evaluations. confidence: high relevance: high limitations: - Evaluation is limited to LibriTTS test-clean (English, read speech), leaving performance on spontaneous speech, noisy environments, and non-English languages uncharacterized. The SECS speaker similarity scores for the proposed models fall below the strongest baseline (Lee et al. 2024), indicating room for improvement in speaker faithfulness despite strong subjective SMOS scores. The paper does not release code, limiting reproducibility and adoption. RTF comparisons are not fully apples-to-apples since baselines generate complete utterances in batch mode while SMAM operates incrementally. Future directions mentioned include a fully streaming pipeline covering codec processing and applying depthwise decoding strategies to decoder-only speech language models. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1098 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS architecture: - transformer-enc-dec relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - transformer_encoder_decoder claims: - claim_id: bert_prosody_prediction_extendable_to_multi_speaker_without_accent_labels role: supports claim: BERT-based prosody predictors for pitch-accent languages can be extended to multi-speaker settings through speaker style embedding injection without requiring accent label annotation. source: §2.1, §3.4, Table 2, Table 3 evidence: GST-BERT-TTS reduces log f0 MSE from 0.773 (fo-BERT) to 0.266 and raises top-level accent correctness from 47.5% to 65% on Hi-Fi-CAPTAIN. confidence: high relevance: high - claim_id: speaker_style_embeddings_generalise_poorly_to_unseen_speakers role: complicates claim: Speaker style embeddings derived from seen-speaker averages do not generalise well to unseen speakers in label-free prosody prediction. source: §4.1 evidence: The authors report that using the GST embedding of an unseen speaker during inference produced flat, monotonous prosody, attributed to insufficient speaker diversity in the training corpus. confidence: medium relevance: low - claim_id: prosody_accuracy_gains_do_not_automatically_raise_mos role: complicates claim: Improvements in prosody prediction accuracy and accent correctness do not automatically translate into higher overall MOS in multi-speaker TTS. source: §3.4, Figure 3 evidence: GST-BERT-TTS outperforms fo-BERT on MSE and accent correctness, yet its MOS remains lower than TTS+Accent, attributed partly to vocoder artefacts on a high-variability corpus and to MOS conflating naturalness with prosodic correctness. confidence: high relevance: high - claim_id: multitask_prosody_prediction_synergistic_with_speaker_conditioning role: supports claim: Multi-task prosody prediction covering fundamental frequency, energy, and duration jointly improves over pitch-only prediction when speaker conditioning is present. source: §3.4, Table 2 evidence: Adding energy and duration prediction to fo-BERT (speaker-agnostic) had minimal impact on log f0 MSE (0.773 vs. 0.776), whereas GST-BERT (speaker-conditioned) showed a synergistic improvement, reducing log f0 MSE from 0.302 to 0.266. confidence: medium relevance: low limitations: - Evaluation uses a proprietary internal Japanese corpus; the accent correctness listening test involves only two speakers, limiting generalisation to other languages or larger speaker sets. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1115 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS architecture: - diffusion - autoregressive-LM relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - autoregressive_generation - diffusion_generation claims: - claim_id: unified_multimodal_emotion_encoder_enables_flexible_control role: supports claim: Unified multi-modal emotion encoders that align different prompt modalities to a shared emotion representation space enable flexible emotion control in zero-shot TTS without sacrificing speaker similarity. source: §2.2, §3.4, Table 1 evidence: Removing the MPEE module drops MOS from 3.73 to 3.20 and emotion accuracy from 48% to 35% in speech-prompt mode, while SSMOS (3.73) is maintained even when emotion and timbre are drawn from different speakers. confidence: high relevance: high - claim_id: emotion_consistency_loss_improves_prosody_emotion_alignment role: supports claim: Auxiliary emotion consistency losses applied to autoregressive prosody prediction improve emotion alignment in generated speech. source: §2.3, §3.4, Tables 1-3 evidence: Removing ECL reduces ESMOS from 4.05 to 3.83 and ACC from 48% to 45% in speech-prompt mode, with a larger effect for text and image prompts. confidence: medium relevance: high - claim_id: emotion_accuracy_remains_below_ground_truth_despite_conditioning role: complicates claim: Emotion accuracy in expressive zero-shot TTS remains substantially below ground truth even with explicit emotion conditioning and auxiliary training objectives. source: §3.4, Tables 1-3 evidence: Best system ACC is 48% versus 54% for ground-truth mel-spectrograms, a 6-point gap that persists despite MPEE and ECL. confidence: high relevance: low - claim_id: baseline_comparison_limited_by_absence_of_official_implementation role: complicates claim: Evaluating multi-modal TTS systems against prior work is complicated by the absence of official open-source implementations for key baselines. source: §3.4 evidence: MM-TTS has no official open-source release; all comparisons in this paper are against the authors' own reproduction, which the authors acknowledge as a limitation of the experimental evaluation. confidence: high relevance: low - claim_id: hierarchical_disentanglement_enables_independent_timbre_emotion_control role: supports claim: Hierarchical disentanglement of speech attributes at different granularity levels enables fine-grained independent control over timbre and emotion in zero-shot TTS. source: §2.1, §3.4, Tables 1-3 evidence: SSMOS remains stable (3.73-3.76) across all three prompt modalities despite using different emotion sources than the timbre reference. confidence: medium relevance: high limitations: - The system is fine-tuned on only 36 hours of MEAD-TTS (48 actors, 8 emotions), evaluates only English speech, and WER is noticeably worse than ground-truth reconstructed mel (23.4% vs. 18.8%). caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1122 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS architecture: - GAN - transformer-enc-dec relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - transformer_encoder_decoder - gan_adversarial_generation claims: - claim_id: qat_more_effective_than_channel_reduction_at_equal_size role: supports claim: Quantisation-aware training is more effective than parameter reduction via channel shrinkage for achieving compact TTS models at equivalent model sizes. source: §3.3.1, Table 1 evidence: A channel-reduced 32-bit small model at 12.78 MB collapses to MOS 1.20, while the 1.58-bit quantized model at a comparable 4.39 MB scores MOS 3.09. confidence: high relevance: high - claim_id: vocoder_more_sensitive_to_quantisation_than_acoustic_model role: supports claim: The vocoder stage of a TTS pipeline is more sensitive to aggressive weight quantisation than the acoustic model, making it a disproportionate source of quality degradation. source: §3.3.1 evidence: 4-bit vocoder quantisation (MOS 2.57) performs worse than 1.58-bit vocoder quantisation (MOS 3.18), an inversion the authors attribute to the vocoder's higher sensitivity to precision loss. confidence: medium relevance: high - claim_id: sub_2bit_quantisation_preserves_intelligible_synthesis_selectively_applied role: supports claim: Extremely low-bit (1-2 bit) weight quantisation can preserve intelligible and reasonably natural speech synthesis when applied selectively to convolutional layers. source: §3.3.1, Table 1 evidence: 1.58-bit quantization with weight indexing applied to both the acoustic model and vocoder reaches an 83% size reduction (4.39 MB) at MOS 3.09, substantially above the same-size 32-bit channel-reduced baseline. confidence: high relevance: high - claim_id: hardware_aware_packing_necessary_for_subbyte_storage_savings role: complicates claim: Hardware-aware packing schemes are necessary to realise the theoretical storage savings of sub-byte quantisation on devices that operate in 8-bit units. source: §2.3 evidence: Naively stored ternary weights still occupy 8 bits each on standard hardware; the paper's weight-indexing scheme packs 5 consecutive ternary weights (3^5=243) into a single int8 index to realise the storage benefit. confidence: high relevance: low limitations: - The subjective evaluation uses only 15 raters and 30 utterances; all experiments use a single TTS architecture (JETS with convolutional blocks) trained on a single corpus, so generalisation to transformer-heavy or autoregressive architectures is untested. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1210 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - VC architecture: - diffusion relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - diffusion_generation claims: - claim_id: diffemotionvc_dual_granularity_emotion role: supports claim: Dual-granularity emotion feature extraction (combining utterance-level and frame-level representations) improves emotion discriminability in voice conversion compared to single-scale approaches. source: §2.1.3, §3.3.3, Table 3 evidence: DiffEmotionVC's dual-granularity emotion encoder achieves 80% ECA and 0.78 Pearson Corr on the ESD dataset; ablation confirms removing the emotion encoder is the most damaging intervention, dropping Corr to 0.38. confidence: high relevance: high - claim_id: diffemotionvc_orthogonality_stable_disentanglement role: supports claim: Orthogonality constraints on emotion, speaker, and content feature spaces provide a stable and effective disentanglement mechanism for emotional voice conversion. source: §2.2.2, §3.3.3, Table 3 evidence: Removing orthogonal loss reduces SECS from 0.73 to 0.70 and Corr from 0.78 to 0.72; the paper explicitly motivates orthogonal loss as a remedy for the training instability of the mutual information loss used in prior work. confidence: high relevance: low - claim_id: diffemotionvc_arousal_confusion role: complicates claim: Diffusion-based EVC systems achieve strong overall emotion accuracy but struggle to discriminate between high-arousal emotions sharing similar arousal-valence profiles. source: §3.3.1 evidence: DiffEmotionVC reaches 80% ECA overall but the paper notes difficulty distinguishing happy, surprised, and angry, attributing this to insufficient emotional diversity in the ESD training data rather than a fundamental model limitation. confidence: medium relevance: low - claim_id: diffemotionvc_discretization_leakage role: complicates claim: Discretisation of continuous speech representations degrades emotion voice conversion by introducing content-emotion feature leakage. source: §3.3.2, Table 2 evidence: Replacing continuous ContentVec with VQ-ContentVec drops UTMOS from 4.04 to 2.54 and Corr from 0.78 to 0.50; SpeechTokenizer RVQ1 discrete features produce the worst performance (UTMOS 1.79), demonstrating that discrete tokens cause timbre and emotion entanglement. confidence: high relevance: high - claim_id: diffemotionvc_crossattention_fusion role: supports claim: Cross-attention fusion outperforms additive fusion for integrating heterogeneous speech features in voice conversion systems. source: §3.3.3, Table 3 evidence: Ablation replacing gated cross-attention with simple additive fusion reduces UTMOS from 4.04 to 3.26, a 19% degradation in predicted audio quality. confidence: high relevance: high limitations: - Evaluation is limited to the ESD dataset (five emotions, primarily Mandarin Chinese; the ablation table specifically targets the zh-Angry subset), restricting generalisability to other languages and broader emotion categories; model size is unreported and no code is released. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1229 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - VC architecture: - flow-matching relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - flow_matching_generation claims: - claim_id: icl_flow_matching_enables_joint_timbre_and_background_preservation role: supports claim: In-context learning with flow matching can enable voice conversion systems to simultaneously transform speaker timbre and preserve background sounds without an explicit separation step. source: §4.2, Table 2 evidence: E2E-BPVC achieves BS-MOS 4.60 and SS-MOS 4.02, comparable to the cascaded Denoise-VC II (BS-MOS 4.65, SS-MOS 4.07), using a single model without a denoising module, validated by 12 human raters. confidence: high relevance: high - claim_id: standard_vc_metrics_insufficient_for_background_preserving_systems role: complicates claim: Standard voice conversion evaluation metrics (speaker similarity, character error rate, speech quality) are insufficient for assessing systems that operate on speech with background sounds. source: §4.1, Table 1 evidence: ECAPA-TDNN speaker similarity and ASR-based CER are degraded by background sound, causing clean-output systems to appear relatively stronger on objective metrics despite failing entirely on background preservation (ICL-VC BS-MOS 0.70); the authors explicitly state objective metrics 'do not adequately reflect the capabilities' of background-preserving systems. confidence: high relevance: high - claim_id: noise_robust_ssl_representations_improve_content_disentanglement role: supports claim: Noise-robust self-supervised speech representations improve content disentanglement in voice conversion systems trained on speech with background sounds. source: §4.3, Table 3 evidence: Replacing HuBERT with WavLM as the semantic token backbone reduces CER from 10.27 to 7.99 under noisy evaluation, and training k-means on noisy speech further reduces CER to 7.99 versus 9.22. confidence: high relevance: low - claim_id: background_preservation_objective_trades_off_against_clean_speech_quality role: complicates claim: Designing a single voice conversion model to handle background preservation introduces a competing objective that slightly degrades clean-speech conversion quality relative to a clean-speech-only system. source: §4.1, Table 1 evidence: E2E-BPVC achieves SIM 0.849 and CER 3.18 on clean-speech conversion, modestly below the clean-only ICL-VC (SIM 0.873, CER 2.37). confidence: medium relevance: low limitations: - Evaluation relies exclusively on LibriTTS with synthetically added noise and music at controlled SNRs, without testing on real-world recordings with natural, non-stationary backgrounds; the comparison set is limited to two baselines sharing the same ICL-VC foundation. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1364 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - singing - TTS architecture: - diffusion relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - diffusion_generation claims: - claim_id: schrodinger_bridge_consistency_training_improves_one_step_quality role: supports claim: Integrating Schrödinger bridge into consistency training can improve one-step generative model quality for audio synthesis without a pre-trained teacher model. source: §2.2, §3.4, Table 2 evidence: VS-Singer's CSB decoder achieves MOS 3.74 at NFE=4 on test-unseen, outperforming teacher-distilled CoMoSpeech (3.70); ablation removing the Schrödinger bridge drops MOS from 3.81 to 3.50 and raises MCD from 7.65 to 8.63. confidence: high relevance: high - claim_id: cross_modal_visual_attention_encodes_spatial_acoustics_for_singing role: supports claim: Cross-modal visual attention over scene images can encode spatial acoustic properties into a singing synthesis pipeline, enabling reverberation-aware binaural audio generation. source: §2.1, §3.4, Table 2 evidence: Removing the Modal Interaction Network degrades LRE from 0.889 to 1.257 and RTE from 0.058 to 0.084 in ablation on test-seen. confidence: medium relevance: low - claim_id: one_step_consistency_inference_incurs_quality_penalty role: complicates claim: One-step consistency model inference incurs a substantial quality penalty compared to multi-step inference in diffusion-based singing synthesis. source: §3.3, Table 1 evidence: VS-Singer at NFE=1 achieves MOS 3.48 vs. 3.74 at NFE=4 on test-unseen, with MCD degrading from 7.71 to 8.55; CoMoSpeech shows the same pattern. confidence: high relevance: high - claim_id: unified_e2e_spatial_audio_outperforms_cascaded_pipelines role: supports claim: Unified end-to-end architectures for spatial audio generation outperform cascaded specialised model pipelines in both inference speed and spatial accuracy. source: §3.3, Table 1 evidence: VS-Singer (NFE=1) achieves RTF 0.024 on test-unseen vs. RTF 0.203 for a cascaded DiffSinger baseline, while achieving better LRE and RTE than all cascaded combinations at NFE=4. confidence: medium relevance: low limitations: - The full system is evaluated on a single female Chinese pop singer with synthetic room-impulse-response binaural training data; no code is released, and perceptual evaluation uses only naturalness MOS without spatial audio perceptual tests. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1394 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS architecture: - transformer-enc-dec relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - transformer_encoder_decoder claims: - claim_id: diemotts_selfsupervised_distillation_vs_grl_vq role: supports claim: Self-supervised distillation with emotion-specific inductive biases can learn speaker-independent emotion embeddings more effectively than GRL-based or VQ-based disentanglement. source: §3.4, Table 1 evidence: DiEmo-TTS surpasses Trans-GRL, Trans-VQ, and Trans-Ort on eMOS across all four emotion categories in subjective evaluation, without requiring explicit speaker labels during emotion encoder training. confidence: high relevance: high - claim_id: diemotts_formant_perturbation_disentangle role: supports claim: Formant-based speaker perturbation is more effective for disentangling speaker identity from emotion than non-targeted noise augmentation in cross-speaker emotion transfer. source: §3.6, Table 2 evidence: Replacing formant perturbation with MUSAN/RIR noise augmentation in the ablation increases WER and degrades SECS, whereas formant perturbation exploits the timbre-formant correlation to distort identity while preserving emotional expression. confidence: high relevance: low - claim_id: diemotts_dct_conditioning_balance role: supports claim: Multi-factor conditioning via shared attention mechanisms produces better balance between speaker fidelity and emotional expressiveness than concatenation conditioning in transformer-based TTS. source: §3.6, Table 2 evidence: Adding the DCT block improves eMOS from 3.89 to 4.07 while keeping sMOS comparable in the ablation; the block applies weight-sharing multi-head attention per style signal and fuses outputs via MLP. confidence: high relevance: high - claim_id: diemotts_ssl_labeled_dependency role: complicates claim: Achieving speaker-independent emotion representations via self-supervised methods still depends on labeled auxiliary data for emotion space definition. source: §4, §3.2 evidence: The emotion clustering step requires a WavLM-based emotional attribute predictor fine-tuned on the MSP-Podcast labeled corpus; the authors identify dependence on this predictor as a limitation for generalising to unseen speakers. confidence: medium relevance: low limitations: - Evaluation uses only 2 target speakers (one male, one female from ESD), both trained exclusively on neutral utterances; WER of 16.16% indicates non-trivial intelligibility degradation, and scalability to languages beyond English is undemonstrated. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1397 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - singing - VC architecture: - diffusion relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - diffusion_generation claims: - claim_id: vibesvc_dwt_f0_decomposition_control role: supports claim: Explicit frequency-domain decomposition of the F0 contour enables more accurate and controllable singing style transfer than implicit style-embedding approaches. source: §4.1.1, Table 1 evidence: VibE-SVC achieves 0.700 style accuracy in style-only conversion versus 0.213-0.525 for SoVITS baselines using direct style embeddings, on the VocalSet straight/vibrato benchmark. confidence: high relevance: low - claim_id: vibesvc_mpd_style_accuracy role: supports claim: Adversarial training on the target frequency band of the F0 contour improves singing style accuracy without degrading naturalness. source: §4.3, Table 1 evidence: Removing the multi-period discriminator reduces style accuracy from 0.700 to 0.625 and MOS from 4.124 to 4.016 in the style-only conversion experiment. confidence: high relevance: high - claim_id: vibesvc_naturalness_tradeoff role: complicates claim: Increasing style transfer accuracy in singing voice conversion trades off against naturalness, and explicit disentanglement does not fully eliminate this tension. source: §4.1.2, Figure 3 evidence: Figure 3 shows a consistent inverse correlation between MOS and style accuracy across all baselines and VibE-SVC; the highest-accuracy model (VibE-SVC) has lower naturalness than the highest-naturalness baseline (SoVITS with style embedding, 0.213 style accuracy). confidence: medium relevance: high - claim_id: vibesvc_dwt_level_sensitivity role: refines claim: The effective granularity of F0-based singing style disentanglement via DWT is sensitive to decomposition level, with an optimal level that captures vibrato without including unrelated high-frequency content. source: §4.3, Table 3 evidence: Style accuracy is 0.163 at DWT level 3 (vibrato information absent), 0.694 at level 4 (optimal), and drops slightly at level 5 due to inclusion of irrelevant high-frequency components. confidence: high relevance: low - claim_id: vibesvc_continuous_vibrato_control role: supports claim: Vibrato extent in singing voice conversion can be controlled continuously at inference time by scalar multiplication of the isolated high-frequency F0 component, without retraining. source: §4.2, Table 2, Figure 5 evidence: Scaling the high-frequency F0 contour from 0.1 to 2.0 produces style accuracy ranging from 0.066 to 0.928; frame-level control is also demonstrated by applying scaling at specific target frame indices. confidence: high relevance: low limitations: - The model is restricted to two singing styles (straight and vibrato) derived from VocalSet and does not address falsetto, breathy voice, belt, or melisma; human evaluation uses at minimum 20 Amazon MTurk raters per model, on the low end for resolving the small MOS differences reported. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1434 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - VC architecture: - diffusion relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - diffusion_generation claims: - claim_id: rewind_time_reversal_augmentation role: supports claim: Full-utterance time reversal can serve as an effective signal-level data augmentation for speaker representation learning in voice conversion, as it suppresses phonemic content while retaining speaker-discriminative tonal features. source: §3.1, Table 1 evidence: A perceptual study shows 80.3% speaker identification accuracy from time-reversed speech; Table 1 confirms complete reversal achieves 100% WER (full linguistic removal) alongside the highest cosine speaker similarity score (0.96), higher than any short-time reversal window. confidence: high relevance: high - claim_id: rewind_embedding_fusion_similarity role: supports claim: Fusing speaker embeddings from augmented training signals with conventional embeddings improves speaker similarity in zero-shot diffusion-based voice conversion. source: §4.3, Table 2 evidence: Adding reversed-speech speaker embeddings via a weighted fusion layer (α = β = 0.5) improves objective speaker similarity by 4.16% on average across DiffHierVC and DDDM-VC; DDDM-VC objective SPK-SIM rises from 0.70 to 0.79 and subjective MUSHRA from 77.46 to 78.61. confidence: high relevance: high - claim_id: rewind_backbone_variance role: complicates claim: The effectiveness of speaker embedding augmentation in voice conversion varies substantially across backbone architectures, complicating claims of generalisability. source: §4.3, Table 2 evidence: For DiffVC, the augmentation improves subjective speaker similarity (50.12 to 53.42) but reduces objective similarity (0.75 to 0.71), while DiffHierVC shows objective improvement but negligible subjective change; only DDDM-VC shows consistent gains on both metrics. confidence: medium relevance: high - claim_id: rewind_no_quality_tradeoff role: supports claim: Improving speaker disentanglement through augmentation does not necessarily trade off against generated speech quality in diffusion-based voice conversion. source: §4.3, Table 2 evidence: DDDM-VC+Ours improves both WV-MOS (3.84 to 3.91) and UTMOS (3.21 to 3.55) alongside speaker similarity gains, indicating that stronger speaker conditioning from the STR augmentation does not degrade synthesis quality. confidence: high relevance: high limitations: - The perceptual study supporting the time-reversal principle is small (25 participants, 6 speakers, English only); the approach has been evaluated only on diffusion-based VC systems, and no code is publicly released. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1478 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS architecture: - hybrid relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - hybrid_generation claims: - claim_id: mobile_video_encoders_reduce_lip_to_speech_complexity_at_intelligibility_cost role: supports claim: Replacing ResNet-18/AV-HuBERT visual encoders with mobile video networks reduces lip-to-speech system complexity by an order of magnitude at the cost of intelligibility degradation. source: §4.1, Table 1 evidence: LightL2S uses MoViNet-A0 instead of ResNet-18-based AV-HuBERT, reducing inference from 32-34 GMacs to 0.8 GMacs while WER increases from 27-30% (AV-HuBERT-based methods) to 64.8%. confidence: high relevance: low - claim_id: multires_spectrogram_discriminator_improves_ddsp_naturalness role: supports claim: Multi-resolution spectrogram discriminators substantially improve speech naturalness in DDSP-based lip-to-speech synthesis, beyond spectral regression alone. source: §4.3, Table 3 evidence: Removing the adversarial GAN loss collapses UTMOS from 2.93 to 1.58 and SECS from 0.72 to 0.63, with computational cost unchanged. confidence: high relevance: high - claim_id: objective_metrics_diverge_from_human_judgment_in_lip_to_speech role: complicates claim: Objective quality metrics can diverge from human perceptual judgments in lip-to-speech synthesis, making WER and UTMOS insufficient as sole quality signals. source: §4.2, Table 2 evidence: NaturalL2S exceeds ground truth on UTMOS (3.66 vs. 3.59) but scores lower in subjective naturalness MOS (4.10 vs. 4.51); LightL2S achieves the highest speaker similarity MOS despite being outranked on SECS by several baselines. confidence: high relevance: high - claim_id: zipformer_substitutes_conformer_with_quality_and_cost_gains role: refines claim: Efficient transformer variants (Zipformer) can substitute standard Conformer backbones in visual speech modelling with simultaneous improvements in quality and computational cost. source: §4.3, Table 3 evidence: Replacing Zipformer with a Conformer backbone increases GMacs from 0.80 to 1.09 while reducing UTMOS from 2.93 to 2.74 and raising WER from 64.8% to 66.3%. confidence: medium relevance: high limitations: - Intelligibility remains a major limitation, with WER 64.8% more than double AV-HuBERT-based methods; evaluation is confined to English TED/TEDx video, and parameter count is not reported. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1531 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - singing - VC architecture: - VAE - GAN relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - gan_adversarial_generation - variational_latent_generation claims: - claim_id: svc_dimreduction_timbre_leakage role: supports claim: In singing voice conversion, reducing the dimensionality of SSL embeddings through random channel selection proportionally reduces timbre leakage while preserving sufficient phonetic content for accurate reconstruction. source: §4. Results, Table 1 evidence: SSL-128-Emb (128 of 768 HuBERT dimensions) achieves SMOS 3.920 vs. SSL-Emb's 2.750 on the Chinese test set, while maintaining CMOS 4.244 vs. 4.366 for full embeddings. The pattern holds for SSL-Soft and ContentVec variants. confidence: high relevance: high - claim_id: svc_discrete_token_generalization_failure role: complicates claim: Discrete token-based content representations for voice conversion cannot generalize to phonetic inventories outside the training distribution. source: §4. Results, Table 2 evidence: SSL-Token CMOS drops from 4.086 on Chinese to 3.210 on other languages (English, Korean, Vietnamese, Japanese, Cantonese), demonstrating that k-means quantization of HuBERT with up to 10,000 clusters still loses phonetic detail that is language-specific. Continuous dimension-reduced embeddings maintain CMOS above 4.17 across both conditions. confidence: high relevance: high - claim_id: svc_ssl_dimension_uniform_distribution role: supports claim: Self-supervised speech embedding dimensions contain roughly proportional timbre and content signal, such that uniform random subsampling functions as an effective form of timbre disentanglement. source: §2.2, §4. Results evidence: The uniform-distribution assumption underlying random dimension selection is empirically supported by consistent improvements across three distinct SSL embedding types (HuBERT, SSL-Soft, ContentVec), each responding to dimension reduction in the same direction. confidence: medium relevance: low - claim_id: svc_unsupervised_matches_supervised_disentangle role: refines claim: Supervised disentanglement methods for SSL content encoders in voice conversion can be matched or exceeded by unsupervised dimensionality reduction at equivalent embedding sizes. source: §4. Results, Table 1, Table 2 evidence: SSL-256-Emb surpasses SSL-Soft-Emb (which uses supervised soft target training) at the same 256-dimensional size on both SSIM and SMOS metrics in Chinese. ContentVEC-256-Emb matches token-based singer similarity while outperforming token-based CMOS on cross-lingual evaluation. confidence: medium relevance: high limitations: - All training data is proprietary (200h internal Chinese singing corpus); no public benchmark is used, and the optimal dimension count d is determined by grid search with no principled selection criterion. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1550 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS - VC - evaluation architecture: - transformer-enc-dec - VAE - GAN relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - transformer_encoder_decoder - gan_adversarial_generation - variational_latent_generation claims: - claim_id: diacritization_improves_arabic_tts_intelligibility role: supports claim: Diacritization of Arabic text input significantly improves TTS intelligibility and subjective quality in Modern Standard Arabic synthesis. source: §4.1, Table 4 evidence: VITS trained with diacritics achieved 20.67% WER versus 35.69% without, and was the only model with a statistically significant preference for the diacritized variant in pairwise preference tests. confidence: high relevance: high - claim_id: multispeaker_training_improves_perspeaker_quality role: supports claim: Multi-speaker training improves synthesis quality for individual target speakers relative to speaker-specific training on equivalent data volume. source: §4.1, Figure 2b evidence: Multi-speaker VITS was strongly preferred over single-speaker VITS trained on the same speaker's 2.69 hours in a pairwise preference test. confidence: medium relevance: high - claim_id: synthetic_augmentation_improves_arabic_tts_intelligibility role: supports claim: Synthetic speech data augmentation can improve TTS intelligibility even when the augmenting speech is of variable quality. source: §4.1 evidence: Training VITS on all ArVoice parts including synthetic speech (Google Standard and WaveNet TTS) reduced WER by 8.2% absolute versus human-speech-only training. confidence: medium relevance: high - claim_id: multilingual_llm_tts_fails_without_arabic_adaptation role: complicates claim: Multilingual LLM-based TTS systems do not reliably transfer to Arabic without adequate language-specific training data or adaptation, and can produce unintelligible output despite fine-tuning. source: §4.1, Table 4 evidence: Fish-Speech, a dual-autoregressive LLM-based multilingual system, achieved WER above 100% on Arabic after fine-tuning on ArVoice human speech. confidence: high relevance: high limitations: - The professionally recorded portions require a Data Usage Agreement, human speech totals only ~10 hours across 7 speakers, and preference tests use only 25 samples per pair. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1625 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - VC architecture: - GAN relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - gan_adversarial_generation claims: - claim_id: mimic_blocker_defense_preserves_perceptual_quality role: supports claim: Active adversarial defenses can obstruct voice-conversion speaker extraction while preserving the perceived identity and quality of protected audio. source: §3.2, Table 1 evidence: Mimic Blocker reports PESQ 3.633 and STOI 0.942 with attack success above 0.88, while its listening test found 99.7% same-speaker judgments for protected versus original audio. confidence: medium relevance: high - claim_id: mimic_blocker_ssl_targets_transfer_across_vc role: supports claim: Self-supervised speech representations can provide model-agnostic attack targets that transfer active defenses across voice-conversion architectures. source: §3.1, §3.2, Table 1, Table 2 evidence: The same WavLM- or HuBERT-targeted defense reached attack success near 0.88 on FreeVC and 0.99 on TriAAN-VC without retraining. confidence: high relevance: low - claim_id: mimic_blocker_waveform_perturbation_quality role: supports claim: Waveform-domain adversarial perturbation can preserve perceptual naturalness better than spectrogram-domain defenses that require vocoder reconstruction. source: §2.2, §3.2, §3.3, Table 1 evidence: The waveform-domain method reached PESQ 3.633 versus 1.61–1.99 for prior spectrogram-domain baselines, alongside 99.7% same-speaker listener judgments. confidence: medium relevance: high - claim_id: mimic_blocker_no_target_speaker_training role: refines claim: Active voice-conversion defense need not require target-speaker training data when perturbations maximize distance in pretrained speech-representation space. source: §2.3, §1 evidence: Training uses only the protected speaker's speech and separates original and perturbed SSL embeddings instead of relying on style-target speech pairs. confidence: high relevance: low limitations: - The listening panel contained only 6 evaluators assessing 50 sets, limiting statistical robustness and evaluator diversity. - Evaluation covers two voice-conversion models and one English corpus, so perceptual findings may not generalize. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1638 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS architecture: - hybrid relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - hybrid_generation claims: - claim_id: eatsspeech_disentangled_emotion_branch role: supports claim: Treating emotion as a disentangled parallel synthesis branch improves emotional expressiveness over global style conditioning in zero-shot TTS. source: §3.6, Table 2 evidence: Ablation removing emotion decoupling reduces EMOS from 3.96 to 3.45 on LibriTTS test, with Emotion Discrepancy rising from 0.571 to 0.939, confirming that explicit emotion isolation drives the expressiveness gains. confidence: high relevance: high - claim_id: eatsspeech_llm_emotion_transform role: supports claim: LLM-based emotion transformation conditioned on target text semantics reduces emotion-content mismatch in zero-shot speech synthesis. source: §3.6, Table 2 evidence: Removing the emotion-aware LLM raises ED from 0.571 to 1.056 and drops EMOS from 3.96 to 3.73 on LibriTTS test, with both components contributing independently to emotional consistency. confidence: high relevance: high - claim_id: eatsspeech_direct_transfer_mismatch role: complicates claim: Direct reference emotion transfer in zero-shot TTS produces emotionally inconsistent speech when the reference and target texts differ in emotional valence. source: §3.3, §3.5, Table 1 evidence: All five comparison zero-shot TTS baselines (YourTTS, TransferTTS, VALL-E, E2-TTS, CosyVoice) show Emotion Discrepancy scores of 0.802 to 1.399 versus EATS-Speech at 0.571, suggesting that without explicit emotion adaptation, standard conditioning mechanisms misalign emotion to content. confidence: medium relevance: low - claim_id: eatsspeech_circular_metric_evaluation role: complicates claim: Evaluation of emotion expressiveness with metrics derived from the same model used in training introduces circularity that limits the reliability of reported Emotion Discrepancy scores. source: §2.1, §3.5 evidence: ED is computed using utterance-level emotion2vec embeddings, the same pre-trained model whose frame-level features the Emotion Tokenizer builds on; this means the metric and the system share the same representational basis, potentially inflating reported gains. confidence: medium relevance: low limitations: - Evaluation is limited to LibriTTS, an audiobook corpus with naturally constrained emotional variability; model size is not reported, and only English is evaluated. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1639 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - codec - VC architecture: - GAN relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - gan_adversarial_generation claims: - claim_id: lombardtokenizer_rvq_layer_distillation role: supports claim: Targeted distillation into specific RVQ layers can isolate fine-grained speaking style attributes (such as vocal effort) independently of semantic content in neural speech codecs. source: §2.3, §3.1, Table 2 evidence: LombardTokenizer conditions the second RVQ layer via cosine distillation from Lombard speech encoders while keeping RVQ layer 1 semantically focused via mHuBERT; the resulting system achieves WER 10.97% and EER 6.67% on vocal effort conversion on AVID, substantially outperforming FreeVC (WER 20.35%, EER 16.67%) while retaining comparable synthesis quality. confidence: high relevance: low - claim_id: lombardtokenizer_disentangle_quality_tradeoff role: complicates claim: Disentanglement constraints in neural codecs reduce unconstrained reconstruction quality relative to codecs without regularisation. source: §3.1, Table 1 evidence: EnCodec (no disentanglement) achieves PESQ 3.32 and STOI 0.94 on LibriSpeech, while LombardTokenizer's dual distillation (semantic and Lombard) yields PESQ 3.07 and STOI 0.93, consistent with SpeechTokenizer's 3.01; the quality gap widens on in-distribution neutral speech but narrows on expressive zero-shot data. confidence: high relevance: low - claim_id: lombardtokenizer_encoder_integration_matters role: complicates claim: Providing a specialized style encoder to a voice conversion model architecture does not guarantee that the encoder's information will be effectively exploited for style control. source: §3.2, Table 2, Figure 2 evidence: FreeVC modified with the same Lombard encoder (FVClmb) fails to produce statistically distinguishable vocal effort levels in perceptual evaluation, and degrades WER on FLombard (32.63%) relative to the standard speaker-encoder variant (24.04%), while LT1 using the same encoder via RVQ distillation achieves WER 17.74% and accurate perceptual effort transfer. confidence: high relevance: high - claim_id: lombardtokenizer_crosslingual_generalization role: supports claim: Zero-shot generalization of speaking style transfer across languages is achievable when codec disentanglement is guided by multilingual self-supervised representations. source: §2.1, §3.2, Table 2 evidence: LombardTokenizer trained on English AVID achieves WER 15.77% on the unseen French FLombard dataset (zero-shot), compared to FreeVC's 24.04-32.63%; the paper attributes part of this advantage to replacing HuBERT with multilingual mHuBERT in the semantic RVQ layer. confidence: medium relevance: low - claim_id: lombardtokenizer_crossspeaker_low_leakage role: supports claim: Codec-level disentanglement enables robust cross-speaker style transfer with low speaker identity leakage under both intra-speaker and inter-speaker conditions. source: §3.2, Table 2, Figure 2 evidence: LombardTokenizer inter-speaker vocal effort conversion produces perceptual distributions not significantly different from intra-speaker conversion (Dunn's test), with EER remaining below 8% on AVID for both LT1 (6.67%) and LT2 (7.83%), indicating preserved speaker identity. confidence: high relevance: high limitations: - The zero-shot evaluation is from English training to French test data, and both languages are Indo-European; the AVID dataset uses instructed intensity levels rather than naturally elicited Lombard speech, which may reduce ecological validity. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1684 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS architecture: - hybrid relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - hybrid_generation claims: - claim_id: saras_speaker_aware_retrieval role: supports claim: Incorporating speaker identity into the retrieval query improves style compatibility in RAG-based zero-shot TTS. source: §3.5, Table 1 evidence: Ablation removing the speaker encoder from Zero-shot CLAP drops S-SMOS from 3.556 to 3.168, while the full model outperforms CA-CLAP (3.117) on the same metric, confirming that speaker conditioning directs retrieval toward each speaker's stylistic range. confidence: high relevance: high - claim_id: saras_style_timbre_disentangle_retrieval role: supports claim: Style-timbre disentanglement enables retrieval in a style-specific embedding space that better captures expressive variation than entangled audio representations. source: §2.1, §3.3, §3.4 evidence: Split-TTS uses Seed-VC to generate timbre-free NTF features for training, yielding a GST encoder whose embeddings cluster by emotion in t-SNE; this style encoder, used in Zero-shot CLAP retrieval, outperforms the CA-CLAP audio encoder on style-similarity MOS (3.556 vs 3.117). confidence: high relevance: high - claim_id: saras_style_fidelity_tradeoff role: complicates claim: Optimizing for style-compatible reference retrieval in zero-shot TTS can trade off against acoustic fidelity metrics. source: §3.3, Table 1 evidence: SA-RAS achieves better S-SMOS and ST-MOS than CA-CLAP but slightly higher MCD (6.86 vs 6.678), suggesting that the reference best matched for style is not always the one that minimises acoustic distortion. confidence: medium relevance: high - claim_id: saras_oracle_gap role: complicates claim: Retrieval-augmented zero-shot TTS consistently falls short of oracle (text-matched) reference selection by a substantial margin across both naturalness and style metrics. source: §3.4, Table 1 evidence: The proposed SA-RAS scores NMOS 3.827 and S-SMOS 3.556, versus the "Self" oracle at NMOS 4.046 and S-SMOS 3.980, a gap that persists even after speaker-aware retrieval improvements. confidence: medium relevance: high limitations: - The subjective evaluation involves only 15 listeners rating 15 samples drawn from three unseen speakers; the paper does not compare against end-to-end zero-shot TTS systems beyond the retrieval method comparison. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1726 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS - evaluation architecture: - hybrid relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: - hybrid_generation claims: - claim_id: minimal_readaloud_finetuning_matches_hourscale_spontaneous role: supports claim: Fine-tuning large-scale TTS models on minimal scripted read-aloud data (order of minutes) produces personalised voices with perceived AAC suitability comparable to fine-tuning on hour-scale spontaneous recordings. source: §3.2, Table 1 evidence: FT-Read (12 min of read-aloud sentences, SECS 0.643) and FT-Spont-Short (1h spontaneous, SECS 0.706) show no significant difference in MUSHRA-rated listener preference (p=0.213). confidence: high relevance: high - claim_id: dysarthric_zeroshot_prompt_matches_typical_aac_suitability role: supports claim: Zero-shot TTS adaptation from dysarthric reference audio achieves perceptual suitability for assistive communication comparable to adaptation from typical speech, enabling personalisation for users who lack pre-morbid recordings. source: §3.2 evidence: ZS-Atypical and ZS-Typical show no significant difference in MUSHRA-rated AAC suitability (p=0.683), despite a large objective speaker similarity gap (SECS 0.324 vs. 0.635). confidence: high relevance: high - claim_id: secs_may_not_predict_aac_perceived_suitability role: refines claim: Embedding-based objective speaker similarity metrics may not reliably predict perceived suitability in assistive communication contexts, where other voice qualities matter more than timbre fidelity. source: §3.1, §3.2, Table 1 evidence: ZS-Atypical achieves the lowest SECS (0.324) of all evaluated systems yet is rated comparably to ZS-Typical (SECS 0.635) by human listeners in the MUSHRA-like AAC suitability test (p=0.683). confidence: high relevance: high - claim_id: dysarthric_prompt_yields_measurable_prosodic_alignment role: supports claim: Using dysarthric speech as a prosodic reference input for a fine-tuned TTS system yields measurable alignment with the speaker's intended prosodic pattern while preserving intelligibility. source: §5.2 evidence: Dysarthric-prompted synthesis achieved 29.3% higher SSIM in scalogram comparison of prosodic peaks versus unprompted synthesis (0.0548 vs. 0.0424, p=0.0013); CER improved slightly (2.9% to 2.3%). confidence: medium relevance: low limitations: - The entire evaluation rests on a single research participant (one Hungarian stroke survivor); generalisability across speakers, languages, and impairment types is untested. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1747 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - VC architecture: - diffusion - GAN relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - gan_adversarial_generation - diffusion_generation claims: - claim_id: conversion_process_distillation_avoids_identity_collapse role: supports claim: Performing knowledge distillation in the conversion process rather than the reconstruction process is necessary to avoid identity-mapping collapse when the content encoder is simultaneously trainable. source: §4.2, Table 1, Table 2 evidence: Baseline reconstruction-distillation with trainable CNN achieves UTMOS 3.45 and SECS 0.718; switching to conversion-based ADCD raises these to 4.03 and 0.853. confidence: high relevance: high - claim_id: joint_diffusion_content_encoder_distillation_cuts_cost role: supports claim: Simultaneously distilling both the diffusion backbone and the content encoder of a one-step diffusion VC model can reduce inference cost substantially with little or no quality degradation relative to distilling only the diffusion module. source: §4.3, Table 3, Table 5 evidence: FasterVoiceGrad achieves UTMOS 4.03 and SECS 0.853 vs. FastVoiceGrad's 3.96 and 0.847, while delivering 6.6x GPU RTF and 1.8x CPU RTF improvement. confidence: high relevance: high - claim_id: secs_may_not_reflect_perceptual_speaker_identity_in_vc role: complicates claim: Neural speaker-encoder-based similarity metrics may not reflect perceptual speaker identity as judged by human listeners, particularly when source speaker residuals remain after conversion. source: §4.3, Table 3, Table 4 evidence: FasterVoiceGrad scores higher SECS (0.853 vs. 0.847) than FastVoiceGrad but lower sMOS (2.66 vs. 2.84); the authors attribute the discrepancy to residual source-speaker characteristics. confidence: high relevance: high - claim_id: inverse_score_distillation_mitigates_content_speaker_tradeoff role: refines claim: Contrastive speaker emphasis via inverse score distillation mitigates the trade-off between content preservation and speaker conversion when joint distillation is used. source: §4.2, Table 1 evidence: Adding inverse score distillation on top of reconversion-based distillation improves SECS from 0.844 to 0.853 without degrading CER (1.2) or UTMOS (4.03). confidence: medium relevance: high limitations: - Subjective evaluation involved only 11 participants and 90 speaker/sentence pairs; the study does not evaluate against modern flow-matching or large-scale VC baselines. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1763 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - VC architecture: - GAN - diffusion relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - gan_adversarial_generation - diffusion_generation claims: - claim_id: partial_vocoder_upsampling_sufficient_for_adversarial_discrimination role: supports claim: Intermediate vocoder features after partial upsampling are sufficient for waveform-quality adversarial discrimination of acoustic feature generators. source: §3.2, Tables 1 and 5, Figure 3 evidence: FVG + VPFD1 (single upsampling step, 8x) matches full waveform discriminator performance on UTMOS, DNSMOS, CER, and SECS across both VCTK and LibriTTS. confidence: high relevance: high - claim_id: frozen_pretrained_feature_extractor_essential_for_projected_gan role: supports claim: Freezing pretrained feature extractors in projected GAN discriminators is essential for acoustic synthesis quality. source: §3.2, Table 2 evidence: Ablation shows that both pretraining and freezing V_feat are independently necessary; omitting either degrades UTMOS, DNSMOS, and SECS. confidence: high relevance: high - claim_id: waveform_discriminators_effective_but_resourceheavy role: complicates claim: Waveform-domain discriminators in two-stage TTS/VC training are effective but impose resource costs that make them impractical outside well-resourced settings. source: §1, §3.3, Tables 1 and 3 evidence: VWD requires 47 hours and 66.3 GB GPU memory on VCTK; mel-spectrogram discriminators avoid this cost but fail to match waveform-domain quality on DNSMOS. confidence: high relevance: high - claim_id: minimum_one_upsampling_step_needed_for_periodic_structure role: refines claim: For vocoder-based feature projection, a minimum of one upsampling step is necessary to produce the periodic structures required for effective adversarial discrimination. source: §3.2, Table 1, Figure 3 evidence: VPFD0 (no upsampling) significantly degrades DNSMOS (3.66 vs. 3.79) and SECS (0.843 vs. 0.851) relative to VPFD1. confidence: medium relevance: high limitations: - VPFD is validated only on one VC system (FastVoiceGrad on VoiceGrad) and never on a TTS system, despite TTS being a stated target application; subjective evaluation involves only 11 participants. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1776 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS architecture: - hybrid relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - hybrid_generation claims: - claim_id: multi_task_joint_training_on_synthesis_editing_and role: supports claim: Multi-task joint training on synthesis, editing, and continuation tasks improves speech synthesis quality over single-task training in non-autoregressive codec models. source: §3.2, §3.3, Tables 1, 3 evidence: In all three tokenizer configurations (ST, STDAC, HuDAC), multi-task SpeechSEC consistently outperforms the corresponding single-task baseline on MOS, voice preservation, WER, and CER, with gains confirmed by ablation. confidence: high relevance: high - claim_id: in_multi_task_speech_generation_training_editing_tasks role: refines claim: In multi-task speech generation training, editing tasks primarily contribute intelligibility improvements while continuation tasks primarily contribute acoustic quality and voice preservation. source: §3.3, Table 3 evidence: Ablation removing the editing task increases WER by up to 4.4 points with minimal audio quality change; removing continuation degrades MOS by up to 0.18 and voice preservation by up to 0.06 with smaller intelligibility effects. confidence: high relevance: high - claim_id: non_autoregressive_masked_token_prediction_frameworks_can_unify role: supports claim: Non-autoregressive masked token prediction frameworks can unify speech synthesis, editing, and continuation tasks through task-specific input conditioning within a single model. source: §2, §3.2, Table 2 evidence: SpeechSEC handles all three tasks with a shared Conformer backbone, differentiating tasks via a Task Register embedding and per-task masking strategies, achieving competitive quality on editing (MOS 3.93) and continuation (MOS 3.63) alongside synthesis. confidence: high relevance: high - claim_id: the_choice_of_semantic_and_acoustic_token_extractor role: complicates claim: The choice of semantic and acoustic token extractor significantly affects absolute synthesis quality in codec-based TTS, even when model architecture and training are held constant. source: §3.2, Table 1 evidence: With the same SpeechSEC architecture and training scheme, MOS ranges from 3.65 (STDAC) to 4.20 (ST) across the three tokenizer configurations, and voice preservation from 0.61 to 0.72, indicating that tokenizer quality is a dominant factor. confidence: high relevance: high limitations: - The evaluation is restricted to LibriTTS-R, a clean studio-quality English corpus, leaving generalization to noisy, spontaneous, or multilingual speech untested. Cross-system comparisons with SoundStorm use independently reported numbers from separate evaluations, weakening the claim of surpassing prior state of the art. Model parameter count is not reported, preventing meaningful comparisons of capacity-normalised performance. Speech continuation lacks intelligibility metrics (WER, CER) by design, limiting interpretability of those results. The paper does not evaluate the editing task on real-world editing scenarios beyond random masking. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1819 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS architecture: - GAN relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - gan_adversarial_generation claims: - claim_id: input_side_multistream_decomposition_improves_vocoder_quality_efficiency role: supports claim: Input-side multi-stream decomposition via analysis filters can improve both quality and efficiency in harmonic-prior 2D convolutional vocoders. source: §2.2, §4.2 evidence: MS-Wavehax achieves the highest MOS in both analysis-synthesis and TTS tasks while requiring roughly half the MACs of standard Wavehax and only 0.332M parameters. confidence: high relevance: high - claim_id: batch_rtf_poor_proxy_for_streaming_throughput role: supports claim: Batch-inference real-time factor is a poor proxy for streaming throughput; architecture choices interact with chunk size and cache behaviour in ways that can reverse the efficiency ranking. source: §3.2, Table 2, Figure 2 evidence: At small chunk sizes MS-Wavehax achieves the lowest RTF among all vocoders; at larger chunk sizes Vocos overtakes it because 1D filter matrices are cache-resident while Wavehax's 2D im2col matrices grow with chunk size. confidence: high relevance: low - claim_id: multistream_decomposition_not_universally_beneficial role: complicates claim: Multi-stream decomposition does not uniformly improve quality across all frame-level vocoder architectures. source: §4.2 evidence: MS-Vocos underperforms standard Vocos in MOS evaluation, attributed to a latent representation narrower than the output complex spectrum. confidence: medium relevance: high - claim_id: batchnorm_enables_streaming_and_quality_gains role: supports claim: Replacing layer normalization with batch normalization in GAN vocoders enables streaming compatibility and can also yield objective quality improvements. source: §2.3 evidence: Switching from layer normalization (incompatible with incremental streaming) to batch normalization improved PESQ and UTMOS scores in internal experiments across all tested configurations. confidence: medium relevance: high - claim_id: causal_vocoder_matches_noncausal_quality_singlespeaker role: supports claim: Causal vocoder processing with minimal lookahead can match non-causal quality in single-speaker TTS settings. source: §4.2, Figure 4 evidence: Causal MS-Wavehax with a one-frame lookahead achieves nearly identical MOS to non-causal MS-Wavehax on the JSUT TTS task, whereas the gap persists in the multi-speaker A/S task. confidence: medium relevance: high limitations: - Evaluations are conducted exclusively on Japanese speech corpora on a single CPU architecture; no end-to-end latency measurement with an acoustic model frontend is included. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-1940 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS architecture: - transformer-enc-dec - flow-matching relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - transformer_encoder_decoder - flow_matching_generation claims: - claim_id: stochastic_prosody_predictors_achieve_naturalness_parity_with_diversity role: supports claim: Stochastic prosody predictors achieve naturalness parity with human recordings while enabling prosodic diversity that deterministic predictors cannot produce. source: §3.4, Figure 5, Figure 6 evidence: RF at 0.4 temperature shows no statistically significant difference from human naturalness ratings; the deterministic baseline achieves highest naturalness but lowest diversity. confidence: high relevance: high - claim_id: cascaded_prediction_outperforms_joint_for_duration role: refines claim: For explicit prosody modeling, cascaded prediction of prosodic features outperforms joint prediction for duration but the ordering of pitch and energy within the cascade has negligible impact. source: §3.3, Table 1 evidence: Joint prediction yields 17% higher duration JS divergence (0.562) vs. cascading (0.481), while pitch-first and energy-first orderings differ by less than 0.005. confidence: medium relevance: low - claim_id: naturalness_diversity_inversely_correlated role: complicates claim: Naturalness and prosodic diversity are inversely correlated, constraining simultaneous optimization of both properties in TTS. source: §3.4, Figure 5, Figure 6 evidence: No system achieves both maximal naturalness and maximal diversity; even human speech exhibits this trade-off, scoring highest on diversity but not on naturalness. confidence: medium relevance: high - claim_id: sampling_temperature_enables_runtime_prosody_control role: supports claim: Sampling temperature in generative prosody predictors enables effective runtime control over prosodic variability without retraining. source: §3.3, Figure 2, Figure 3 evidence: Pitch contour variance and duration variance increase monotonically with temperature for all stochastic methods, following a near-exponential relationship. confidence: high relevance: low limitations: - Experiments are restricted to read speech (LibriTTS); the stochastic models do not yet capture the full diversity of human prosodic distributions, remaining narrower and more unimodal than RAVDESS reference distributions. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-2031 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS architecture: - transformer-enc-dec relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - transformer_encoder_decoder claims: - claim_id: phonotactic_rule_assignment_improves_crossfamily_zeroshot_tts role: supports claim: For zero-shot TTS across language families, assigning G2P rules by phonotactic family rather than script family produces substantially more natural and intelligible synthesis. source: §5.3, Tables 1, 2 evidence: Sanskrit synthesis with Dravidian rules (Kannada system, MOS 4.12, WER 4.6%) substantially outperforms synthesis with Indo-Aryan rules (Hindi system, MOS 3.25, WER 28.0%). confidence: high relevance: high - claim_id: extended_shared_phoneme_inventory_enables_unseen_phone_zeroshot role: supports claim: A shared phoneme inventory extended via phonetic substitution can enable zero-shot synthesis for target languages with sounds absent from all source training sets. source: §4.2 evidence: The CLS was extended to represent nukta-diacritic consonants and borrowed aspiration in Kurukh by mapping them to nearest-neighbour base characters, enabling intelligible synthesis without Kurukh training data. confidence: medium relevance: low - claim_id: dialect_zeroshot_tts_requires_phonotactic_source_matching role: supports claim: Dialect-specific zero-shot TTS requires matching the source synthesiser to the dialect's phonotactic affiliation, not its surface script or nominal language grouping. source: §5.3, Table 1 evidence: Maharashtrian Konkani (IA phonotactics) is best with Marathi (MOS 3.64) while Canara Konkani (DR phonotactics) is best with Kannada (MOS 3.34), despite both dialects sharing a language label. confidence: high relevance: high - claim_id: crossfamily_synthesis_degrades_for_structurally_absent_features role: complicates claim: Cross-family zero-shot synthesis degrades for target languages with phonological features structurally absent from all available source synthesisers. source: §5.3 evidence: Kurukh synthesis struggles specifically with glottal stops across both Hindi and Kannada systems, and question intonation was misrendered as assertions in some utterances. confidence: medium relevance: low - claim_id: phonotactic_family_predicts_transfer_better_than_script role: refines claim: Linguistic proximity for zero-shot TTS transfer is captured more precisely by phonotactic family than by script family or geographic proximity. source: §5.3, Tables 1, 3 evidence: 'Sanskrit belongs to the IA family but achieves lower MOS with IA synthesisers (3.25) than with DR synthesisers (4.12), and MCD scores independently confirm the same pattern (IA: 8.48 vs. DR: 7.15/6.94).' confidence: high relevance: high limitations: - The evaluation relies on only four monolingual source synthesisers, and the anomalous MCD result for Maharashtrian Konkani (higher distortion with Marathi despite better MOS) is unresolved. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-2032 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS architecture: - hybrid relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: - hybrid_generation claims: - claim_id: labelconditioned_tokenization_enables_wordstress_control role: supports claim: Explicit prosody conditioning via label-encoded tokenization enables word-level stress control in flow-based TTS without modifying the core model architecture. source: §4.1, Table 1 evidence: GlowEx TTS, encoding stressed words as uppercase tokens in an expanded character lookup table, achieves 88.76% true positive rate under supervised DNN stress detection and outperforms baseline GlowTTS under unsupervised K-Means (83.54% vs. 66.6%). confidence: high relevance: low - claim_id: duration_modification_more_reliable_stress_cue_than_energy role: supports claim: Duration modification is a more reliable cue for perceptible word-level stress exaggeration than energy modification in cascaded TTS-vocoder systems. source: §6.1, §6.2, Table 1 evidence: Duration-based WORLD vocoder modifications consistently achieve higher stress detection true positive rates and higher exaggeration quality MOS than energy-only modifications across all six TTS systems evaluated. confidence: high relevance: high - claim_id: unsupervised_supervised_stress_detection_disagree role: complicates claim: Unsupervised and supervised evaluation methods for prosody detection may disagree on the effectiveness of explicit stress conditioning approaches. source: §6.1, Table 1 evidence: GlowEx TTS outperforms baseline GlowTTS under unsupervised K-Means stress detection (83.54% vs. 66.6%) but shows nearly identical performance under supervised DNN classification (88.76% vs. 89.84%). confidence: high relevance: low - claim_id: energy_modification_reduces_naturalness_despite_stress_gain role: complicates claim: Energy modifications in post-hoc vocoder-based exaggeration can reduce perceived naturalness even while boosting stress prominence, creating a trade-off for applications that require both. source: §6.2, Figure 4 evidence: Energy modifications consistently lower overall quality MOS relative to duration-only modifications across all cascaded TTS systems. confidence: medium relevance: high limitations: - The study is restricted to English with a small, author-annotated 1000-sentence dataset, and the perceptual study recruits speech-processing experts rather than actual L2 learners. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-2151 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - VC - evaluation architecture: - GAN relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - gan_adversarial_generation claims: - claim_id: parallel_corpus_prerequisite_for_objective_vc_metrics role: supports claim: Objective voice conversion evaluation metrics that require ground-truth utterances (MCD, F0-RMSE, FFE, GPE) are only feasible when a parallel corpus exists, making dataset parallelism a prerequisite for rigorous VC benchmarking. source: §1.3 evidence: The authors identify the absence of parallel Farsi data as the direct reason these objective metrics cannot be applied to existing Farsi corpora; FaVC is designed specifically to make them feasible. confidence: high relevance: low - claim_id: ganvc_on_lowresource_parallel_corpus_matches_english_benchmarks role: supports claim: GAN-based voice conversion systems trained on parallel corpora in low-resource languages can reach naturalness and speaker similarity scores comparable to training on established English benchmarks. source: §4.3, Table 2 evidence: CycleGAN-VC, StarGAN-VC, StarGAN-VC2, and StarGANv2-VC trained on FaVC achieve MOS-N of 3.3-3.94 and MOS-P of 2.5-4.1, within the range of the same models trained on VCC2018. confidence: medium relevance: high - claim_id: nonenglish_vc_limited_by_lack_of_parallel_corpora role: complicates claim: Non-English voice conversion research is limited by the near-absence of parallel, phonemically balanced corpora outside English. source: §1.2, §2.2 evidence: Despite Farsi being spoken by over 100 million people, no publicly available parallel Farsi VC dataset existed prior to FaVC. confidence: high relevance: low - claim_id: film_adain_decoder_conditioning_supports_effective_oneshot_vc role: supports claim: Incorporating FiLM and AdaIN conditioning into a TTS decoder supports effective one-shot voice conversion with results competitive with many-to-many GAN baselines. source: §4.1, §4.3, Table 2 evidence: Modified StyleTTS-VC with FiLM+AdaIN achieves MOS-N 3.94 and MOS-P 3.87 on FaVC, the highest naturalness score among the five evaluated models. confidence: medium relevance: high limitations: - The dataset is small (11 speakers, 405 utterances each); cross-language comparisons conflate language, speaker population, and recording condition differences, and no statistical significance testing is reported for MOS comparisons. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-2189 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS architecture: - transformer-enc-dec relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - transformer_encoder_decoder claims: - claim_id: joint_acoustictextual_masking_beats_style_emotion_encoders role: supports claim: Combining acoustic context with text conditioning in a masked prediction framework produces more accurate prosody prediction than style encoders or emotion representations adapted to the task. source: §4.1, Table 1 evidence: ProMode outperforms StyleTTS2*, Wav2Vec2-SER*, and Emotion2Vec* across all F0 and energy metrics at frame and phoneme level on GigaSpeech Test. confidence: high relevance: low - claim_id: ser_representations_dont_transfer_to_prosody_prediction role: complicates claim: Self-supervised representations trained for emotion recognition do not transfer effectively to prosody prediction, even after full fine-tuning on prosody-labelled data. source: §4.1, Table 1 evidence: Emotion2Vec* and Wav2Vec2-SER* show markedly lower F0 RPA and higher RMSE than StyleTTS2* and ProMode. confidence: medium relevance: low - claim_id: dualdecoder_acousticonly_loss_prevents_encoder_collapse role: supports claim: A dual-decoder architecture with an auxiliary acoustic-only loss prevents the prosody encoder from collapsing onto text context in masked prediction, where removing the loss causes catastrophic degradation. source: §4.3, Table 1 evidence: Ablation removing AOL drops F0 RPA from 43.9% to 25.8%, the largest degradation of any ablation. confidence: high relevance: low - claim_id: taskagnostic_prosody_prediction_improves_downstream_tts role: supports claim: Task-agnostic prosody models that predict F0 independently can improve both naturalness and prosody preference in downstream TTS systems without requiring joint end-to-end training. source: §4.2, Table 2, §4.2.2 evidence: Replacing FluentSpeech's built-in pitch predictor with ProMode-predicted pitch improves UTMOS from 3.00 to 3.10, WER from 4.41% to 3.99%, and AutoPCP from 2.47 to 2.64. confidence: high relevance: high limitations: - All experiments are conducted on English GigaSpeech; ProMode requires ground-truth phoneme durations from a forced aligner during evaluation, and model size is not reported. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-2283 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: pronunciation_metrics_correlate_better_with_accent_quality role: supports claim: Pronunciation-based objective metrics correlate more reliably with accent quality rankings than general intelligibility or naturalness proxies. source: §5.2, Table 2 evidence: PPG CosSim and PPG JS both achieved SRCC=0.9643 with hypothesised accent ranking (p=0.0005); VF RMSE achieved SRCC=0.9286 (p=0.0025). WER's SRCC was 0.6429 and not statistically significant (p=0.1194). confidence: high relevance: high - claim_id: wer_unreliable_for_underrepresented_accent_evaluation role: contradicts claim: WER is a reliable quality metric for evaluating TTS systems on underrepresented accents. source: §5.2, Table 2 evidence: WER ranked copysyn worse than xtts and xtts worse than corrupt30k, despite the hypothesised reverse quality ordering; WER's correlation with the quality ranking was non-significant (p=0.1194), attributed to ASR accent bias. confidence: high relevance: low - claim_id: transcriptions_improve_xab_statistical_efficiency role: supports claim: Providing reference transcriptions to listeners during accent similarity evaluation significantly improves statistical discriminability with fewer participants. source: §5.1, Figure 1 evidence: Baseline XAB with 15 listeners failed to reach p<0.05; XAB+trans+screen achieved significance within 10 valid submissions; the full design reached 64.1±6.7% preference for copysyn. confidence: high relevance: high - claim_id: highlight_tasks_improve_efficiency_but_double_completion_time role: complicates claim: Auxiliary highlight tasks in listening tests improve statistical efficiency but nearly double completion time, limiting benefit to settings with constrained listener pools. source: §5.1 evidence: XAB+trans+highlight took ~30.6 minutes vs. ~16.9 minutes for baseline XAB. confidence: medium relevance: high limitations: - The study evaluates a single accent variety (Edinburgh Scottish English) using two VCTK speakers; the high SRCC for speaker-similarity (WavLM) may be confounded by the corruption scheme affecting broader speaker modelling. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-2573 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS - evaluation architecture: - flow-matching - GAN relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - gan_adversarial_generation - flow_matching_generation claims: - claim_id: g2p_accuracy_critical_for_tts_quality_beyond_spectrogram_metrics role: supports claim: Accurate grapheme-to-phoneme conversion is a critical prerequisite for TTS quality in morphologically complex languages, with phonemizer errors directly degrading perceived naturalness and pronunciation. source: §4.2, Table 2 evidence: The modified espeak-ng phonemizer reduced mispronounced words by 38% (157 to 97) and improved naturalness MOS by 0.17 points for Egyptian dialect; objective MCD scores were unchanged, confirming the perceptual significance of phonemizer accuracy beyond spectrogram-level metrics. confidence: high relevance: high - claim_id: diacritization_essential_for_intelligible_arabic_tts role: supports claim: Diacritization (vowelization) of Arabic text is essential for intelligible TTS; training without vowels causes consonant clustering and robotic-sounding output. source: §4.3, Table 3 evidence: MSA naturalness MOS dropped from 3.21 (vowelized) to 2.67 (unvowelized) and pronunciation from 3.95 to 2.53, with listening revealing consonant clustering artifacts. confidence: high relevance: high - claim_id: code_switching_harder_than_monolingual_tts role: complicates claim: Code-switching TTS is more difficult than monolingual TTS even when both use the same acoustic model and vocoder trained on equivalent data quantities. source: §4.2, Table 2 evidence: CS pronunciation MOS (4.23 with modified phonemizer) is lower than Egyptian (4.57) and MSA (4.27) despite identical training duration and model architecture. confidence: medium relevance: high - claim_id: same_speaker_multidialect_recording_enables_disentanglement role: supports claim: Recording the same speaker across multiple dialects and an L2 language enables controlled disentanglement of speaker identity from dialect-specific acoustic characteristics in TTS datasets. source: §2.2 evidence: SawtArabi records a native Egyptian speaker in EGY, MSA, English, and CS conditions, allowing comparison of L2-speaking characteristics against native dialect production within a constant speaker identity. confidence: medium relevance: low limitations: - The corpus size (4 hours total, 1 hour per condition) is small relative to modern TTS training norms, and baselines have not been validated against multi-speaker or zero-shot TTS settings. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-2586 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS architecture: - transformer-enc-dec - GAN relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - transformer_encoder_decoder - gan_adversarial_generation claims: - claim_id: spotlighttts_voiced_aware_quantization role: supports claim: Restricting style quantization to acoustically informative (voiced) speech regions improves both style expressiveness and prosodic accuracy in reference-based TTS. source: §4.4.1, Table 4.3 evidence: On ESD, removing voiced extraction raises RMSE_F0 from 8.27 to 11.48 Hz and WER from 12.64 to 14.06; the full model achieves the best style similarity MOS (3.84) over all baselines. confidence: high relevance: high - claim_id: spotlighttts_sd_sp_loss_pairing role: supports claim: Pairing style disentanglement with a complementary prosody-preserving loss stabilises training and prevents prosody degradation from aggressive content removal. source: §4.4.2, Table 4.3 evidence: Removing only the SP loss raises pitch error to 9.74 Hz; removing both SD and SP degrades nMOS to 3.66 and WER to 15.38, with the SP-only-removed condition showing worse prosody than the case with no disentanglement losses at all. confidence: high relevance: high - claim_id: spotlighttts_biased_attention_unvoiced_fill role: supports claim: Asymmetric (biased) self-attention in unvoiced region filling, allowing information flow from voiced to unvoiced positions but not the reverse, outperforms symmetric or fully blocked alternatives for prosodic continuity. source: §4.4.3, Table 4.4 evidence: Standard self-attention in the UF module degrades pitch error to 16.38 Hz and F1 v/uv to 0.6668; binary masking partially recovers at 13.19 Hz versus 8.27 Hz for biased attention. confidence: high relevance: low - claim_id: spotlighttts_narrow_corpus_generalization role: complicates claim: Style transfer quality gains from region-specific quantization have been established only within a controlled emotional corpus, leaving open the question of whether the approach generalises to broader speaking styles or modern generative architectures. source: §4.1, Table 4.1 evidence: All baselines are FastSpeech 2 variants evaluated solely on ESD (10 speakers, 5 discrete emotions); no comparison with flow-matching, diffusion, or large-scale systems is provided. confidence: medium relevance: low limitations: - All evaluations are conducted on a single English emotional speech corpus (ESD); baselines are restricted to FastSpeech 2-based systems, and model size is not reported. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-2595 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS - evaluation architecture: [] relevance: high evidence_role: - core_evidence current_role: active_evidence method_family: [] claims: - claim_id: zeroshot_tts_can_match_clinical_stimuli_intelligibility role: supports claim: Zero-shot TTS voice cloning models can replicate clinical speech assessment stimuli with psychometrically comparable intelligibility functions, meeting a necessary condition for their use as synthetic alternatives to recorded corpora. source: §3.1 evidence: Psychometric function thresholds for E2 (0.41 dB above original, p=0.053) and VALL-E X (0.05 dB, p=1.00) did not significantly differ from the original BKB corpus, and slopes were statistically indistinguishable across all three models. confidence: high relevance: low - claim_id: naturalness_quality_dont_guarantee_speaker_identity_replication role: complicates claim: High naturalness and speech quality in zero-shot TTS voice cloning do not guarantee accurate speaker identity replication, revealing a trade-off invisible to standard aggregate quality metrics. source: §3.2, §3.3 evidence: XTTSv2 achieved the highest MOS-X2 ratings and lowest intelligibility thresholds but received significantly fewer 'same speaker' responses; E2 achieved the best speaker replication consistency but lower naturalness, prosody, and social impression scores than XTTSv2. confidence: high relevance: high - claim_id: aggregate_quality_scores_mask_domain_critical_failures role: complicates claim: Aggregate TTS quality scores can mask domain-critical performance failures that only appear at specific signal-to-noise operating points. source: §3.1, §4 evidence: VALL-E X produced comparable psychometric function thresholds to the original BKB corpus but exhibited significant intelligibility deviations above -6 dB SNR, the range most critical for hearing-impaired listeners, a failure that MOS-X2 ratings did not capture. confidence: high relevance: high limitations: - The study evaluates models on a single target speaker with a specific accent (Southern Standard British English male), and results may not generalise across accents or speaking styles; the study population consisted exclusively of normal-hearing listeners. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-2739 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS architecture: - GAN relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - gan_adversarial_generation claims: - claim_id: frequency_domain_filters_suppress_gan_vocoder_artifacts role: supports claim: Frequency-domain learnable filters can simultaneously suppress aliasing and blurring artifacts in GAN vocoders, replacing specialised anti-aliasing upsampling modules. source: §2.2, §3.3, Table 1 evidence: GAFilter, a temporally invariant STFT-domain multiplicative filter, improves all objective metrics over BigVGAN on LibriTTS-dev while allowing removal of the anti-aliasing upsampling-downsampling operations. confidence: high relevance: low - claim_id: frequency_domain_artifact_filtering_generalises_ood role: supports claim: Frequency-domain artifact filtering in GAN vocoders generalises to out-of-distribution speakers and acoustic domains without domain-specific adaptation. source: §3.3, Table 3 evidence: AF-Vocoder, trained only on LibriTTS, outperforms BigVGAN in PESQ and periodicity error on VCTK, LJSpeech, and MUSDB18-HD. confidence: high relevance: low - claim_id: san_training_can_increase_perceptual_artifacts_despite_better_objective_scores role: complicates claim: Slicing adversarial network training frameworks can increase perceptual artifact severity in GAN vocoders despite achieving stronger objective periodicity scores. source: §3.3, Table 1 evidence: AF-Vocoder-SAN achieves the best objective scores on M-STFT (0.7713) and Periodicity (0.0886), but perceptual UTMOS (3.517) is lower than the 3.650 of AF-Vocoder V1 trained with standard least-squares GAN loss. confidence: high relevance: high - claim_id: frequency_domain_bias_improves_gan_training_gradient_stability role: supports claim: Frequency-domain inductive biases improve gradient stability during GAN vocoder training, reducing early-stage gradient variance. source: §3.3, Figure 3 evidence: Gradient curves show lower deviation and faster stabilisation for AF-Vocoder versus BigVGAN in early training steps, attributed to GAFilter filtering out anomalous spectral components. confidence: medium relevance: low limitations: - The subjective evaluation uses only 8 raters and a single MUSDB18-HD test set, which limits the strength of the perceptual claims; inference latency and real-time factor are not reported. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-2765 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: cmos_mushra_overestimate_naturalness_vs_deception_rates role: complicates claim: CMOS and MUSHRA scores can overestimate naturalness relative to human deception rates, because listeners in preference tests match stimuli to reference recordings rather than assess absolute human-likeness. source: §4.1, §4.2, Table 2 evidence: StyleTTS 2 (best open-source system on standard benchmarks) achieves only 50.89% mean HFR against a 74.11% human baseline despite reporting CMOS +0.28 on LJSpeech, and XTTS scores MUSHRA 76.58 (above the 74.78 human reference) yet fools listeners only 41.8% of the time. confidence: high relevance: high - claim_id: low_expressivity_benchmarks_set_artificially_low_evaluation_bar role: complicates claim: Evaluating TTS systems on benchmarks with low expressive variation sets an artificially low bar, as human recordings from such datasets are themselves infrequently identified as human. source: §4.1 evidence: The human reference recordings from LJSpeech/LibriTTS/LibriSpeech score a mean HFR baseline of only 74.11%, compared to 70.68% on the more expressive Expresso benchmark where commercial systems (PlayHT 71.49, ElevenLabs 69.85) match the human rate. confidence: high relevance: high - claim_id: commercial_tts_near_human_deception_rates_expressive_speech role: supports claim: Commercial TTS systems achieve near-human deception rates in zero-shot speaker adaptation to expressive conversational speech, while leading open-source systems remain substantially below this threshold. source: §4.2, Table 2 evidence: On Expresso, PlayHT (HFR 71.49) and ElevenLabs (HFR 69.85) approach the human reference rate (70.68), while the best open-source system F5-TTS reaches only 50.26 HFR, a 20-point gap. confidence: high relevance: high - claim_id: finetuning_narrows_but_doesnt_close_naturalness_gap role: refines claim: Fine-tuning on high-quality expressive data improves but does not close the gap between open-source TTS naturalness and human speech in deception-based evaluations. source: §4.3, Table 3 evidence: Fine-tuning on 40 hours of Expresso improved F5-TTS from 50.26 to 52.22 HFR and VoiceCraft from 30.52 to 43.45 HFR, but neither approached the human level. confidence: high relevance: high - claim_id: digital_voice_quality_and_flat_prosody_drive_synthetic_detection role: supports claim: Digital voice quality artifacts and flat prosody are the primary perceptual cues that allow listeners to reliably identify synthetic speech from open-source systems. source: §4.4, Table 4 evidence: Granular HFR analysis attributes synthetic-speech detection primarily to digital voice quality (36.1%), unnatural pauses (22.8%), and flat/monotonic delivery (20.6%), markers on which commercial systems perform comparably to human speech. confidence: high relevance: high limitations: - The study evaluates only US-English native listeners on English speech; findings about human deception rates may not generalise to other languages, accents, or listener populations with different familiarity with synthetic speech. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-2787 published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - TTS - evaluation architecture: - autoregressive-LM relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - autoregressive_generation claims: - claim_id: vocal_effort_modulation_improves_intelligibility_in_noise role: supports claim: Vocal effort modulation in zero-shot TTS via audio prompt selection can produce meaningful intelligibility gains in adverse acoustic conditions. source: §2.1, §3.1 evidence: Llasa-synthesized speech at higher vocal effort levels showed up to 30% WER reduction in meeting room and cafeteria noise at low SNRs, with acoustic analysis confirming systematic increases in F0, intensity, and high-frequency spectral energy. confidence: high relevance: low - claim_id: vocal_effort_benefit_is_noise_and_snr_dependent role: complicates claim: The intelligibility benefit of vocal effort adaptation in TTS is noise-type and SNR-dependent, and becomes negligible in moderate-to-favorable acoustic conditions. source: §3.1 evidence: WER differences across effort levels converged above 6 dB SNR in most environments; most environments reached WER < 0.2 at SNR > 12 dB regardless of effort level. confidence: high relevance: high - claim_id: snr_and_noise_type_dominate_intelligibility_prediction role: complicates claim: SNR and noise type dominate intelligibility prediction; speaker identity, vocal effort, and text complexity add minimal predictive variance beyond those two factors. source: §3.2, Table 2 evidence: XGBoost regression (R2 = 0.65) dropped to R2 = 0.20 when SNR was removed and to R2 = 0.40 when noise type was removed, while removing vocal effort or voice type had near-zero effect (R2 = 0.64 in both cases). confidence: high relevance: low - claim_id: asr_wer_underestimates_perceptual_difficulty_of_noisy_speech role: complicates claim: ASR-based WER underestimates the perceptual difficulty of noisy speech, particularly in speech-like masking conditions and at threshold SNR regions. source: §4 evidence: Human listeners (n=35) exhibited non-monotonic WER patterns in cafeteria noise and sharp threshold effects not observed in wav2vec 2.0 ASR, and showed greater sensitivity to meeting room noise; human WER ranged 0.1-1.2 with qualitatively different SNR-response shapes. confidence: high relevance: high limitations: - The human listening experiment is small (n=35) and covers only a subset of the conditions evaluated by ASR, limiting statistical power for the human vs. ASR comparison. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: interspeech-2025-bokkahallisatish25_interspeech published_date: "2025-08-17" entry_date: '2026-07-27' year: 2025 venue: Interspeech task: - SCA - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: interactive_voice_conversion_pairing_surfaces_speaker_dependent_bias role: supports claim: Interactive evaluation tools that pair voice conversion with side-by-side response comparison can surface speaker-dependent biases in speech foundation models that automated benchmarks do not capture. source: §2.2, §2.4 evidence: The platform lets users re-ask questions with voice-converted inputs and directly observe changes in response tone, phrasing, and behavior, positioning human interactive comparison as the primary bias discovery mechanism rather than the automated metrics. confidence: medium relevance: high - claim_id: automated_metrics_insufficient_for_fairness_assessment_alone role: complicates claim: Automated speech evaluation metrics are insufficient on their own for assessing fairness and speaker-dependent behavior in speech foundation models. source: §2.3, §2.4 evidence: Even with speech rate, pitch statistics, sentiment analysis, and semantic similarity available, the authors present human interactive comparison as the primary mode of bias discovery, arguing that quantitative measures miss subtle differences in model behavior across speaker identities. confidence: medium relevance: high - claim_id: existing_benchmarks_treat_speaker_attributes_independently role: supports claim: Existing benchmarks for speech foundation models treat speaker attributes independently, limiting the ability to detect intersectional or compounded identity-based biases. source: §1 evidence: Survey of VoxDialogue, Spoken StereoSet, and VoxEval reveals that each evaluates dimensions such as gender or age in isolation; VoxEval explicitly excludes bias and harmful content analysis as outside its scope. confidence: medium relevance: low limitations: - The platform reports no empirical benchmark results, so claims about its effectiveness at bias discovery remain qualitative and user-dependent, and the voice conversion module's perceptual fidelity is not systematically evaluated. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2508.13028' published_date: "2025-08-18" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - transformer-enc-dec relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - transformer_encoder_decoder claims: - claim_id: sarcasm_detector_feedback_loss_biases_tts_toward_detectable_sarcasm role: supports claim: Feedback loss from a sarcasm classifier can be integrated into TTS training to bias synthesized speech toward detector-recognizable sarcastic prosody. source: §2.2, §4.2, Table 2 evidence: In the speech+text detection condition, the proposed model's output achieves F1 70.1% vs. 68.8% baseline; in the speech-only condition, 63.4% vs. 62.2%, though the same detector family is used for both training feedback and evaluation. confidence: medium relevance: high - claim_id: bimodal_sarcasm_detection_outperforms_audio_only role: supports claim: Bi-modal sarcasm detection that combines acoustic features and text substantially outperforms audio-only detection, suggesting that sarcasm in speech is often semantically encoded and not recoverable from prosody alone. source: §4.1, Table 1 evidence: Bi-modal (speech+text) F1 scores exceed speech-only F1 scores across both baseline and proposed models (70.1%/68.8% vs. 63.4%/62.2%). confidence: medium relevance: low - claim_id: twostage_finetuning_viable_for_lowresource_expressive_tts role: supports claim: Two-stage fine-tuning, from neutral read speech to conversational speech and then to target style, provides a viable data strategy for low-resource expressive speech synthesis. source: §2.3, §3.1 evidence: Training proceeds through LibriTTS pre-training (800k iterations), 6.17-hour sitcom conversational fine-tuning, then MUStARD++ sarcasm fine-tuning with the feedback loss active. confidence: medium relevance: high - claim_id: subjective_sarcasm_perception_confounded_with_text_content role: complicates claim: Subjective sarcasm perception in listening tests is difficult to isolate from the sarcastic content of the text, creating ambiguity in whether listeners respond to prosody or semantics. source: §5, §4.3 evidence: Thirteen listeners rated the proposed system's MOS as significantly higher, with 53% of utterances judged to have stronger sarcasm tendency and 49% preferred overall; all stimuli were drawn from a sarcasm-labelled dataset, which the paper notes may prime listeners toward attributing sarcasm regardless of prosody. confidence: medium relevance: high limitations: - 'The evaluation design has a critical circularity: the same detector architecture trained on the same data distribution is used both as the training feedback signal and as the primary objective metric, so detection-score improvements are expected by construction and cannot be treated as independent evidence.' caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2507.16835' published_date: "2025-08-21" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - SCA - evaluation architecture: [] relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: [] claims: - claim_id: stt_quality_dominates_cascaded_pipeline_performance role: supports claim: In cascaded STT-LLM-TTS pipelines, transcription quality has a larger effect on downstream conversational performance than LLM or TTS component selection. source: §4.5 evidence: Switching from Whisper to Google STT produced larger effect sizes on all quality metrics than swapping LLM versions or TTS providers across 5,000+ production interviews in five configurations. confidence: high relevance: low - claim_id: llm_judge_scores_insufficient_proxy_for_user_satisfaction role: complicates claim: Automated LLM-based quality evaluation scores are insufficient proxies for user satisfaction in production voice-based conversational AI. source: §4.6 evidence: LLM-as-a-Judge metrics for conversational and technical quality showed Pearson correlations below 0.11 with candidate post-interview star ratings, despite high internal correlations among the automated sub-metrics. confidence: high relevance: low - claim_id: tts_selection_has_secondary_but_systematic_effect role: supports claim: TTS component selection has a systematic but secondary effect on user experience in production voice-based conversational AI, smaller than the effects of STT or LLM choices. source: §4.5, Table 3 evidence: Cartesia TTS outperformed OpenAI TTS by 0.1-0.2 points across all LLM-judge metrics; the Google+GPT-4.1+Cartesia configuration achieved the highest user satisfaction (4.53/5) and was significantly better than all other configurations in post-hoc tests. confidence: medium relevance: low - claim_id: user_satisfaction_depends_on_factors_beyond_measured_quality_metrics role: complicates claim: User satisfaction in voice-based conversational AI depends on factors beyond measurable conversational and technical quality metrics. source: §4.6, §5 evidence: The consistently weak correlation (below 0.11) between automated quality scores and user star ratings, combined with a higher correlation for soft skills assessment (0.53), suggests perceived interpersonal qualities influence satisfaction independently of assessed question quality. confidence: high relevance: low limitations: - The five configurations were deployed in non-overlapping time windows rather than concurrently, introducing temporal confounds that prevent clean causal attribution of performance differences to component choice alone; only two TTS options were evaluated, limiting robust TTS isolation. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2508.15931' published_date: "2025-08-21" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - evaluation architecture: - transformer-enc-dec relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - transformer_encoder_decoder claims: - claim_id: qvtad_pairwise_framing_label_sparsity role: supports claim: Pairwise comparison framing reduces annotation subjectivity and label sparsity in perceptual attribute modelling more effectively than scalar labelling alone. source: §2.2, §3.1 evidence: Framing timbre attribute detection as a pairwise relative-strength task allows DAG-based transitive augmentation to expand 6,038 original training pairs to 166,409 samples without additional human annotation. confidence: medium relevance: high - claim_id: qvtad_differential_attention_generalization role: supports claim: Differential attention that subtracts shared information between two representations improves cross-speaker generalisation in timbre attribute discrimination. source: §3.2, Table 1 evidence: The RTSA² module computes two attention maps and subtracts them (scaled by a learnable λ) to suppress shared timbral signal and amplify attribute-specific contrast; removing RTSA² causes a 0.71% drop in unseen-speaker accuracy while marginally improving seen accuracy, confirming it trades absolute-identity fit for cross-speaker generalisation. confidence: high relevance: low - claim_id: qvtad_transitivity_augmentation role: supports claim: Transitivity-based data augmentation over sparse pairwise labels significantly expands training coverage and improves model robustness without requiring additional human annotation. source: §3.1, Table 2 evidence: Removing the DSU/DAG data augmentation hurts seen accuracy by 2.12% and unseen accuracy by 1.44%, confirming the augmented pseudo-labels contribute measurable generalisation gains. confidence: high relevance: high - claim_id: qvtad_acoustic_perceptual_inconsistency role: complicates claim: Acoustic signal features and human perceptual judgements of timbre attributes are systematically inconsistent, limiting the effectiveness of handcrafted feature approaches. source: §2.2, Figure 1 evidence: The paper motivates learned differential attention over handcrafted acoustic features by noting systematic mismatches between raw acoustic measurements and human timbre attribute ratings. confidence: medium relevance: high limitations: - Evaluation is confined to VCTK-RVA (110 clean studio-recorded English speakers) with no downstream generation validation; the FACodec encoder is frozen throughout, so it is unclear whether joint fine-tuning would further improve performance. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2508.17031' published_date: "2025-08-23" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - transformer-enc-dec - GAN relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - transformer_encoder_decoder - gan_adversarial_generation claims: - claim_id: rephrasetts_crossmodal_beats_global_speaker role: supports claim: Cross-modal attention between phoneme and audio representations is more effective than global speaker embeddings for preserving localised prosodic and acoustic context in speech insertion tasks. source: §4.3, Table 3 evidence: Ablation experiments show that removing cross-modal attention and replacing it with a global speaker encoder worsens MCD on dev-clean, confirming fine-grained contextual style transfer outperforms speaker-level global conditioning. confidence: medium relevance: low - claim_id: rephrasetts_nar_variable_length_no_alignment role: supports claim: Non-autoregressive TTS architectures with duration prediction can infer variable-length insertion segments without explicit phoneme-level alignment at inference time. source: §3.1, §3.2 evidence: The FastSpeech2-style variance adaptor predicts phoneme-level duration at inference; MFA-derived alignments are used only for training supervision, enabling variable-length output without alignment at inference time. confidence: medium relevance: low - claim_id: rephrasetts_multiloss_reduces_artefacts role: supports claim: Combining local and global adversarial losses with a triplet-based style matching loss reduces perceptual artefacts in synthesised speech segments that L1 reconstruction loss alone cannot eliminate. source: §3.3, §4.3 evidence: Phase two training adds local/global LSGAN losses plus a triplet-based style matching loss on top of phase one's L1 reconstruction; the ablation in Table 3 shows each component contributes to perceptual quality on dev-clean. confidence: medium relevance: high - claim_id: rephrasetts_local_context_graceful_degradation role: supports claim: Speech insertion quality degrades more gracefully with insertion length when the model conditions directly on local audio context rather than full-utterance speaker embeddings. source: §4.2, Table 1 evidence: On long insertions, RephraseTTS scores 3.97 MOS versus 2.80 for MetaStyleSpeech (which relies on global speaker embeddings), a much smaller degradation from the short-insertion MOS scores of both systems. confidence: medium relevance: high limitations: - Evaluation is limited to clean LibriTTS dev-clean/dev-other read speech; the user study comprises only 15 utterances rated by 6 annotators; SpeechPainter, the closest neural inpainting baseline, is not evaluated quantitatively because its task formulation differs. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2508.17494' published_date: "2025-08-24" entry_date: '2026-07-27' year: 2025 venue: workshop task: - TTS architecture: - hybrid relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - hybrid_generation claims: - claim_id: cascaded_ssml_task_decomposition_outperforms_joint_llm_generation role: supports claim: Cascaded task decomposition, separating structural tag prediction from numerical parameter regression, substantially outperforms joint LLM generation for SSML-based prosody control. source: §4.4, §5.4, Table 4, Table 5 evidence: QwenA (structural prediction) achieves 99.24% F1 for break-tag placement vs. 92.06% F1 for fine-tuned BERT; QwenB (numerical regression) achieves MAE of 0.97% pitch, 1.09% volume, 1.10% rate, and 132.9ms break timing, a 25-40% error reduction over the best few-shot LLM baseline. confidence: high relevance: low - claim_id: promptonly_llms_undergenerate_prosodic_markup_tags role: complicates claim: Prompt-only LLMs (zero-shot and few-shot) systematically under-generate prosodic markup tags relative to gold annotations, and this failure persists across architectures and scales. source: §5.3, Figure 3 evidence: All zero-shot and few-shot prompted LLMs benchmarked under-generate break and prosody tags relative to the gold standard, and few-shot prompting can cause structural collapse, e.g. Llama 3's prosody tagging nearly disappears in few-shot mode. confidence: high relevance: high - claim_id: ssml_prosody_enhancement_yields_perceptual_gains_without_retraining role: supports claim: Prosody enhancement via SSML yields substantial perceptual gains over neutral commercial TTS voices, even when the underlying synthesiser is not retrained. source: §5.1 evidence: Perceptual AB testing with 18 listeners over 30 one-minute pairs shows MOS rising from 3.20 (Azure Henri baseline) to 3.87 with SSML enhancement (p<0.005), with 15 of 18 listeners preferring the enhanced version in over half of comparisons. confidence: high relevance: high - claim_id: normalized_french_prosody_captures_linguistically_meaningful_patterns role: supports claim: French TTS prosody normalised relative to a synthetic baseline captures linguistically meaningful patterns, phrase-final pitch rises, deliberate pacing, without requiring manual annotation. source: §3, Appendix A evidence: Prosodic features (pitch, volume, rate, break duration) are normalised relative to the Azure Henri baseline voice using semitone/LUFS/rate-delta conversions, per the force-alignment and syntagm segmentation pipeline. confidence: medium relevance: high limitations: - The dataset is 14 hours of proprietary French podcasts; generalisation to other French domains, speech styles, or languages is unvalidated, and the perceptual test involved only 18 listeners. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2508.17623' published_date: "2025-08-25" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - SCA - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: sdms_fail_to_sustain_emotional_coherence_across_multiturn_dialogue role: supports claim: Current spoken dialogue systems reliably fail to sustain emotionally coherent responses across multi-turn conversations, with degradation compounding across turns. source: §V, Table V evidence: Multi-turn evaluation shows Moshi's perceptual ERS drops from 0.284 to 0.271 and dGSLM's from 0.296 to 0.291, with categorical averages falling to 0.470 and 0.461 respectively, well below the human baseline of 0.925. confidence: high relevance: high - claim_id: automatic_emotion_metrics_overestimate_coherence_on_atypical_systems role: complicates claim: Automatic continuous and categorical emotion metrics correlate moderately with human perceptual judgements on most SDMs, but overestimate emotional coherence on systems with atypical output characteristics. source: §V, Figure 3, Figure 4 evidence: Moshi and dGSLM are overestimated by automatic continuous/categorical metrics relative to human ratings, showing current automatic emotion metrics do not fully capture human perception of extended emotional coherence. confidence: high relevance: high - claim_id: categorical_and_continuous_emotion_objectives_traded_off role: supports claim: Optimising for categorical emotion appropriateness and optimising for continuous affect trajectory alignment represent distinct objectives that current SDMs trade off against each other rather than jointly satisfying. source: §V, Table IV evidence: 'Freeze-Omni leads categorical-based performance (Avg: 0.738) while GLM-4-Voice leads continuous-based scores (ERS: 0.803) in single-turn settings, with the two metric types not co-ranking systems.' confidence: medium relevance: low - claim_id: synthetic_emotional_speech_sufficient_as_evaluation_stimulus role: supports claim: Synthetic emotional speech generated by controllable TTS systems provides sufficient variety and quality to serve as evaluation stimulus material when real annotated data is scarce. source: §IV-A, Table III evidence: The benchmark's synthetic dataset (500 utterances each of neutral/happy/angry/sad plus 1,251 description prompts, ~7.4 hours) is generated via GPT-4 scripting rendered through CosyVoice, and is used alongside real DailyTalk speech (21.7 hours) as the human upper-bound reference. confidence: medium relevance: high limitations: - Human perceptual evaluation covers only 20 randomly sampled dialogues per system evaluated by 20 annotators; at this sample size, inter-rater agreement statistics are not reported, limiting confidence in the perceptual ERS rankings. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.00675' published_date: "2025-08-31" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - transformer-enc-dec relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - transformer_encoder_decoder claims: - claim_id: speaker_specific_phrasing_behaviour_is_a_substantive_source role: complicates claim: Speaker-specific phrasing behaviour is a substantive source of variance in RP insertion that generic multi-speaker models fail to capture, and modeling it explicitly improves both objective and subjective phrasing quality. source: §5.1.2, Table 3, Table 5 evidence: 'For unseen speakers at inference, the paper proposes a lightweight embedding adapter: a two-linear-layer network trained to map PSVM embeddings from the seen-speaker space to the adapted embedding space used by phrasing models with trainable layers, enabling few-shot adaptation with as few as 5–10 reference utterances per unseen speaker without any fine-tuning of the phrasing model itself.' confidence: high relevance: high - claim_id: phoneme_level_language_models_outperform_subword_level_models role: supports claim: Phoneme-level language models outperform subword-level models on phrase break prediction, even at smaller model sizes, because phoneme representations carry acoustic information more directly relevant to pause insertion than subword tokens. source: §5.1.3, Table 4 evidence: 'On the PLM side, the paper experiments with both subword-level models (BERT, XLNet, RoBERTa, ALBERT, DeBERTaV3 in BASE and LARGE variants) and phoneme-level models: Mixed-Phoneme BERT (MP BERT), which introduces sup-phoneme auxiliary tokens aligned with phoneme tokens, and Phoneme-Level BERT (PL BERT), which uses a phoneme-to-grapheme pre-training objective.' confidence: high relevance: low - claim_id: scaling_subword_plms_from_base_to_large_yields role: supports claim: Scaling subword PLMs from BASE to LARGE yields diminishing returns for phrasing tasks, suggesting a representational ceiling specific to this task modality. source: §5.1.3, Table 4 evidence: Scaling subword PLMs from BASE to LARGE yields only modest F0.5 gains (~0.01–0.009) despite tripling parameter count, suggesting that subword representations approach a ceiling for this task (Table 4, §5.1.3). confidence: high relevance: low - claim_id: pre_trained_speaker_verification_embeddings_capture_prosodic_and role: supports claim: Pre-trained speaker verification embeddings capture prosodic and fluency-related characteristics that transfer to phrasing models via few-shot adaptation without fine-tuning, enabling reasonable generalization to unseen speakers. source: §5.2.2, Table 7 evidence: 'For unseen speakers at inference, the paper proposes a lightweight embedding adapter: a two-linear-layer network trained to map PSVM embeddings from the seen-speaker space to the adapted embedding space used by phrasing models with trainable layers, enabling few-shot adaptation with as few as 5–10 reference utterances per unseen speaker without any fine-tuning of the phrasing model itself.' confidence: high relevance: low - claim_id: f0_5_score_and_naturalness_mos_can_diverge role: supports claim: F0.5 score and naturalness MOS can diverge for phrasing models using different PLMs, indicating that objective phrasing accuracy does not fully predict perceived speech naturalness. source: §5.2.3, Table 8 evidence: Notably, MP BERT-based proposed models achieve competitive or higher MOS than BERT-BASE counterparts for unseen speakers despite lower F0.5 scores, suggesting MP BERT's pause placement better aligns with human naturalness preferences even when it underfits the aggregate distribution (Table 8). confidence: high relevance: high limitations: - Training data is exclusively from LibriTTS-R audiobook readings. The resulting phrasing models are likely miscalibrated for spontaneous speech, conversational TTS, or out-of-domain styles; generalization is explicitly flagged by the authors as an open problem. - 'Additional limitations: the paper cannot disentangle the contributions of phoneme vs. subword information within phoneme-level PLMs (since their pre-training also includes grapheme-level objectives), leaving the mechanism of improvement partially unclear. The embedding adapter for unseen speakers assumes that the mapping from PSVM embeddings to trained embeddings is approximately injective and learnable with a small network — this assumption may not hold for speakers whose acoustic characteristics fall outside the training distribution.' caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.00685' published_date: "2025-08-31" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - autoregressive-LM relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: - autoregressive_generation claims: - claim_id: optimizing_lm_based_tts_with_preference_data_across role: supports claim: Optimizing LM-based TTS with preference data across multiple quality dimensions simultaneously is more effective than per-metric or naively combined ranking approaches. source: §4.5, Table 2 evidence: 'Objective metrics (Table 2): - MPO achieves CER 3.9, SPK-SIM 0.577, prosody (log F0 RMSE) 0.279 vs. SFT baseline CER 4.72, SPK-SIM 0.548, prosody 0.337. - A model trained on combined rankings (naive merge of three per-metric ranks) reaches CER 4.3, SPK-SIM 0.564, prosody 0.218—better on prosody but worse on intelligibility and speaker similarity than MPO. - Single-metric DPO variants confirm that optimizing for one dimension improves that metric but degrades others.' confidence: high relevance: high - claim_id: vanilla_dpo_applied_to_lm_based_tts_without role: supports claim: Vanilla DPO applied to LM-based TTS without regularization causes progressive degradation of generation capability, with CER rising sharply after 10k training steps. source: §4.4, Table 1 evidence: 'Regularization ablation (Table 1): Without the CE constraint, CER worsens from 4.72 (SFT) to 6.41 (10k steps) and 14.52 (15k steps).' confidence: high relevance: low - claim_id: adding_a_cross_entropy_loss_on_preferred_sequences role: supports claim: Adding a cross-entropy loss on preferred sequences as a regularizer during DPO training prevents model collapse without sacrificing alignment gains. source: §3.2, §4.4 evidence: The combined training loss is L = λ·L_DPO + L_CE, where L_CE is the standard cross-entropy objective on the preferred sequences. confidence: high relevance: low - claim_id: multi_dimensional_preference_alignment_in_lm_based_tts role: supports claim: Multi-dimensional preference alignment in LM-based TTS produces output quality comparable to ground truth in subjective pairwise preference tests. source: §4.5, Figure 3 evidence: 'Subjective ABX test (Figure 3): MPO is preferred over the SFT baseline in 52.3% of comparisons (22.4% ties); MPO beats combined-ranking in 40.2% vs. 30.7% (29.1% ties); MPO is competitive with ground truth in pairwise preference.' confidence: high relevance: high limitations: - '- Evaluation is entirely on an internal Mandarin TTS corpus; there is no open benchmark comparison, making it difficult to assess absolute quality relative to published systems. - The model size is not disclosed and the base LM architecture details are partially withheld (internal codec), limiting reproducibility. - The 100-hour preference optimization corpus is Mandarin only; whether the approach generalizes to multilingual or cross-lingual zero-shot scenarios is untested. - The preference set construction requires running three separate quality scorers (ASR, speaker verification, F0) over 10 generated candidates per input—a non-trivial computational overhead not quantified in the paper. - The optimal value of λ (set to 10) and the preference margin thresholds (0.1 for SPK-SIM and prosody) are not ablated; sensitivity to these hyperparameters is unknown. - Comparison against SpeechAlign, UNO, and RIO is discussed only qualitatively; a controlled quantitative comparison against these direct baselines would strengthen the claims.' caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: 2025.clicit-1.27 published_date: "2025-09-01" entry_date: '2026-07-27' year: 2025 venue: workshop task: - TTS - evaluation architecture: - autoregressive-LM relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - autoregressive_generation claims: - claim_id: orthographic_pre_processing_of_input_text_accent_marking role: supports claim: Orthographic pre-processing of input text (accent marking, grapheme substitution, elision resolution) can reduce segmental substitution errors in LLM-based TTS applied to low-resource or historical languages without any model retraining. source: §3.2, §5.2 evidence: Replacing rare digraphs and adapting graphemes (〈c〉 to 〈k〉, 〈ae〉 to 〈ai〉, 〈qu〉 to 〈kw〉) reduced segmental substitutions by approximately one third; no retraining of GPT-4o-mini-tts was performed. confidence: high relevance: low - claim_id: natural_language_prompt_instructions_can_steer_general_purpose role: supports claim: Natural language prompt instructions can steer general-purpose LLM-based TTS toward prosodically-correct speech in low-resource languages without fine-tuning. source: §3.3, §3.4 evidence: A concise system prompt specifying slow pace, explicit stress, and syllable articulation, combined with orthographically pre-processed input, produced at least one expert-validated recording for all 216 Latin verse lines. confidence: high relevance: low - claim_id: llm_based_tts_systems_trained_on_modern_multilingual role: complicates claim: LLM-based TTS systems trained on modern multilingual data produce systematic cross-lingual phonological interference when applied to historically or orthographically distinct languages. source: §5.1 evidence: GPT-4o-mini-tts imposed Italian, English, or Spanish prosodic templates on Latin text depending on vocabulary similarity; the phenomenon confirms that Latin-specific data is essentially absent from the model's training corpus. confidence: high relevance: low - claim_id: prompt_based_prosody_control_is_more_effective_for role: complicates claim: Prompt-based prosody control is more effective for stress placement than for enforcing absolute phonemic length distinctions. source: §5.6, §6 evidence: The workflow approximated Latin vowel quantity through slower pacing on ictic syllables but could not enforce a fixed heavy-to-light duration ratio, the standard for quantitative metre in phonetic work. confidence: high relevance: low limitations: - The corpus depends on GPT-4o-mini-tts, a proprietary model with inaccessible weights. Should the API access policy change, exact reproduction of the dataset becomes impossible. - 'The corpus is small (216 lines, approximately 24 minutes), falling well short of the material typically required for full end-to-end TTS fine-tuning. Stylistic range is narrow: both texts belong to the Augustan literary period and formal register; colloquial Latin, post-Classical spelling, and late-antique forms are untested. A single synthetic voice is used throughout, precluding multi-speaker or gender-diverse analysis. Only the classical reconstructed pronunciation is represented; ecclesiastical pronunciation is excluded. Expert validation, while rigorous, introduces selection bias from a single institution''s phonological tradition. Quantity is approximated through pacing rather than explicit duration control; no fixed heavy-to-light ratio is enforced, leaving a gap relative to the phonetic gold standard.' caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2506.23367' published_date: "2025-09-03" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - flow-matching relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - flow_matching_generation claims: - claim_id: targeted_duration_manipulation_more_effective_than_global_slowing_l2 role: supports claim: Targeted duration manipulation of specific phonemic contrasts is more effective for L2 speech comprehension than undifferentiated global speech rate reduction. source: §3.3.1, §3.3.2, Tables 2, 4 evidence: Clarity mode (1.6x stretch on tense vowels only, 0.75x elsewhere) achieved 15.15% total WER vs. 19.82% for global 1.2x stretch and 24.30% for base in the double-word experiment; tense-vowel WER improved from 60.23% (base) to 29.48% with targeted stretch vs. 37.57% with full stretch in the single-word experiment. confidence: high relevance: low - claim_id: asr_wer_not_reliable_proxy_for_l2_intelligibility role: contradicts claim: ASR-based word error rate can serve as a reliable proxy for L2 listener intelligibility in TTS evaluation. source: §3.3.3, Table 7 evidence: Whisper ASR showed no benefit from clarity mode (17.68% WER vs. 17.10% for base) while L2 listeners achieved their lowest WER with clarity mode (15.15%); ASR's tense/lax substitution pattern (71.42% minimal-pair substitutions on base TTS) differed fundamentally from L2 human error distributions; overall ASR WER varied by less than 3% across conditions that produced large differences in human WER. confidence: high relevance: high - claim_id: global_speech_slowing_not_effective_clarity_strategy_l2 role: complicates claim: Global speech rate reduction is an effective clarity strategy for L2 listener experience. source: §3.3.1, §3.3.2, Tables 2, 3, 5, 6 evidence: Full-stretch TTS scored significantly lower on naturalness (nMOS ~4.2 vs. ~7.0-7.9 for other conditions), prosody, encouragement, and respect in both single- and double-word experiments; L2 participants rated both "too fast" (base) and "too slow" (stretch) as less respectful and encouraging. confidence: high relevance: high - claim_id: l2_subjective_intelligibility_diverges_from_actual_comprehension role: complicates claim: L2 listeners' subjective perceived intelligibility judgments reflect their actual comprehension accuracy for TTS systems. source: §3.3.1, §3.3.2, §5 evidence: Despite clarity mode producing lower objective WER than emphasis in both experiments, L2 participants consistently rated emphasis (stretching all target words) as most intelligible subjectively; participants believed that overall phrase slowing was most intelligible even when it was not, suggesting they were unaware of the duration mechanism improving their comprehension. confidence: high relevance: high limitations: - The study is restricted to French-L1 English-L2 speakers at relatively high proficiency, covering only three English tense/lax vowel pairs; generalisability to other L1 backgrounds and proficiency levels is untested. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.03292' published_date: "2025-09-03" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - evaluation architecture: - hybrid relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: - hybrid_generation claims: - claim_id: ssl_reps_trained_on_natural_audio_generalise_to_synthetic_quality_prediction role: supports claim: Self-supervised audio representations trained on natural audio can generalise to predicting perceptual quality dimensions of synthetic speech and audio without synthetic training examples. source: §IV.C, Table I evidence: The model, trained exclusively on natural audio (AES-natural corpus), achieves SRCC improvements over the challenge baseline across all four perceptual axes on the official AudioMOS 2025 evaluation set of synthetic TTS/TTA/TTM samples, e.g. Production Quality SRCC 0.896 vs. 0.866 baseline. confidence: high relevance: high - claim_id: triplet_loss_improves_rank_correlation_over_mse_only role: supports claim: Triplet loss applied to intermediate embeddings with score-based sampling improves rank correlation with human perceptual ratings over MSE-only training objectives. source: §III, §IV.C, Table II evidence: Adding the triplet loss (L_total = L_MSE + 0.2*L_triplet) with FIFO-buffer anchor/positive/negative sampling raises Content Usefulness SRCC from 0.810 (baseline) to 0.894 and KTAU from 0.629 to 0.724. confidence: high relevance: high - claim_id: multiaxis_quality_prediction_benefits_from_task_specific_heads role: supports claim: Multi-axis perceptual quality prediction benefits from task-specific attention heads over a shared backbone, reflecting the partial independence of production and content quality dimensions. source: §II.C evidence: Each of the four aesthetic axes (Production Quality, Production Complexity, Content Enjoyment, Content Usefulness) is scored by its own multi-head self-attention branch over a shared BEATs-LSTM backbone. confidence: medium relevance: high - claim_id: domain_shift_harms_absolute_score_more_than_rank_ordering role: complicates claim: Domain shift between natural and synthetic audio is more problematic for absolute score prediction than for rank-based ordering of perceptual quality. source: §IV.C, Table II evidence: Content Enjoyment shows the one MSE regression (3.991 vs. baseline 1.142) despite improved rank correlation (SRCC 0.904 vs. 0.841), indicating the model ranks perceptual quality correctly but with larger absolute deviations under domain shift. confidence: high relevance: high limitations: - The model is trained exclusively on 2,700 natural audio samples yet evaluated entirely on synthetic audio; absolute score calibration is unreliable for Content Enjoyment specifically, and no comparison against other published automatic quality predictors (UTMOS, DNSMOS) is included. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.03940' published_date: "2025-09-04" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - SCA - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: paralinguistic_appropriateness_weakest_capability_in_spoken_dialogue_models role: supports claim: Acoustic quality (paralinguistic appropriateness) is the most consistently underperformed capability across current spoken dialogue models, even for proprietary systems, lagging well behind text generation and reasoning dimensions. source: '§Experiments: LLM-based Result, Table 3' evidence: 'Across the LLM-based dimensions, Paralinguistic Appropriateness is universally the weakest dimension (GPT-4o: 3.82, best open-source Step-Audio: 3.43), well below Contextual Coherence (GPT-4o: 4.48).' confidence: medium relevance: high - claim_id: parameter_count_does_not_predict_speech_synthesis_quality_dialogue role: supports claim: 'Model parameter count does not reliably predict speech synthesis quality in spoken dialogue systems: smaller, better-trained models can match or exceed much larger counterparts on naturalness metrics.' source: '§Experiments: Metric-based Results, Table 2' evidence: Step-Audio at 132B achieves the highest BertScore F1 (84.16) but scores poorly on speech naturalness (UTMOS 2.42), while Qwen2.5-Omni at 7B achieves UTMOS 3.57, nearly matching GPT-4o's 3.66. confidence: medium relevance: high - claim_id: acoustically_aware_llm_judge_aligns_with_human_judgment_roleplay role: supports claim: LLM-based evaluation augmented with discrete acoustic features achieves strong alignment with human judgment on multi-dimensional spoken role-playing tasks. source: '§Evaluation Framework, §Experiments: Subjective experimental results' evidence: Human evaluation on 20 sampled instances yields a Pearson correlation of 0.762 between the acoustically-aware LLM judge scores (which augment transcribed text with pitch, energy, speaking-rate bins, and Emotion2Vec labels) and human annotator scores. confidence: medium relevance: high - claim_id: optimal_context_window_length_exists_for_spoken_roleplay role: supports claim: There is an optimal context window length for spoken role-playing models beyond which additional conversational history degrades rather than improves performance. source: '§Experiments: Ablation study on context length, Table 4' evidence: An ablation over context window length (4-10 turns) shows a non-monotonic trend, with a 6-turn context achieving the best overall score. confidence: medium relevance: low limitations: - Human evaluation for judge validation covers only 20 dialogue instances, and character-quality validation covers only 20 characters rated by 5 annotators; the benchmark is sourced exclusively from movie dialogue, which is scripted and stylistically distinct from spontaneous conversation. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.04072' published_date: "2025-09-04" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - autoregressive-LM - flow-matching relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - autoregressive_generation - flow_matching_generation claims: - claim_id: narrative_aware_segmentation_yields_higher_emotional_diversity role: supports claim: Narrative-aware segmentation of audiobook data into character quotation and narration subsets yields training material with measurably higher emotional diversity than standard sentence-level audiobook splits. source: §3.2, Table 5 evidence: The filtered high-expressivity subset (Q_f, 379 hours) retains quotations where Phi-4 extracts a speech adverb or explicit expressive speech verb from narrative context, yielding pseudo-labels such as "he whispered softly." confidence: medium relevance: low - claim_id: flow_matching_gains_more_from_expressive_finetuning_than_ar role: supports claim: Flow-matching TTS models show larger expressivity gains from fine-tuning on targeted expressive speech data than autoregressive models with equivalent training setups. source: §4.3, Table 2 evidence: Fine-tuning F5-TTS (flow-matching) on Q_f raises ContextMOS from 2.95 to 3.33, while fine-tuning SparkTTS (autoregressive) leaves ContextMOS roughly unchanged (2.89-2.97 vs. 2.94 baseline) despite substantial WER improvement. confidence: medium relevance: high - claim_id: narrative_context_conditioning_improves_contextual_appropriateness role: supports claim: Conditioning TTS synthesis on surrounding narrative context rather than only the target utterance text improves contextual appropriateness of synthesized speech at the cost of modest intelligibility degradation. source: §4.3, Table 2 evidence: Training SparkTTS from scratch with contextual narrative conditioning raises ContextMOS from 3.09 to 3.15 but degrades WER from 4.8 to 9.5. confidence: medium relevance: high - claim_id: opensource_tts_less_expressive_than_narrators_despite_comparable_mos role: complicates claim: Current open-source TTS systems are substantially less expressive than human audiobook narrators on contextual benchmarks, even when naturalness scores (MOS) are comparable. source: §5.2, Table 4 evidence: In zero-shot benchmarking, IndexTTS2-Context reaches the best ContextMOS (3.45) and Win-Rate (54%), matching human ground truth, while the other benchmarked systems (SparkTTS, F5-TTS, MaskGCT) show comparable naturalness MOS but negative CMOS (~-0.97 to -0.98) relative to ground truth, unlike IndexTTS2 (+0.25). confidence: medium relevance: high - claim_id: llm_extracted_delivery_pseudolabels_reliable_for_expressive_tts_training role: supports claim: LLM-extracted speech-delivery pseudo-labels (verbs and adverbs) from narrative prose are reliable enough to serve as training signals for expressive TTS, achieving high precision when confidence-filtered. source: §3.3, Figure 2 evidence: The confidence-filtered pseudo-label extraction pipeline (Phi-4 with self-reported confidence scoring) is used to build the 379-hour Q_f subset that drives the largest expressivity gains among all training variants. confidence: medium relevance: low limitations: - ContextMOS and Win-Rate are computed with Gemini-2.5 Pro as judge on only a 201-quotation subset, not the full test set; MOS and CMOS subjective scores use only 30 samples (2 per speaker), limiting statistical precision. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.04093' published_date: "2025-09-04" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS - SCA architecture: - autoregressive-LM - flow-matching relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - autoregressive_generation - flow_matching_generation claims: - claim_id: fullduplex_conversational_data_improves_tts_prosodic_naturalness role: supports claim: Incorporating spontaneous full-duplex conversational data into TTS training improves the prosodic naturalness of synthesized speech, as measured by both acoustic distance metrics and human preference. source: §4.1, §4.2, Table 2 evidence: Fine-tuning CosyVoice-300M on 15 hours of dual-track conversational speech reduced F0 Wasserstein distance by 7.08% (Chinese) and 3.67% (English), and the fine-tuned model was preferred by a majority in A/B preference tests in both languages (45.0% vs. 33.9% Chinese; 46.4% vs. 40.7% English). confidence: medium relevance: high - claim_id: existing_corpora_lack_overlap_interruption_annotations role: supports claim: Existing TTS and dialogue corpora lack the annotated overlap and interruption dynamics needed for full-duplex conversational modelling. source: §1 evidence: Turn-based datasets like Switchboard constrain to sequential exchanges with limited overlap, and DailyTalk rarely includes spontaneous overlapping speech or explicit interruption annotations, leaving a data gap for full-duplex conversational TTS. confidence: medium relevance: high - claim_id: small_corpus_finetuning_gains_inconsistent_across_acoustic_dims role: complicates claim: Fine-tuning TTS models on small conversational corpora yields perceptible naturalness gains but leaves spectral fidelity inconsistent across languages and acoustic feature dimensions. source: §4.1, §4.2, Table 2, Table 3 evidence: MOS naturalness improved by only 0.03-0.04 points in both Chinese and English, and English spectrum l2 distance worsened by 0.85% despite improvements in F0 (3.67%) and ZCR (7.44%), suggesting that small-scale fine-tuning does not uniformly improve all acoustic dimensions. confidence: medium relevance: high limitations: - The datasets are small (10h Chinese / 5h English, 14 unique speakers total), and the validation experiment tests only fine-tuning an existing TTS model rather than end-to-end full-duplex spoken dialogue system improvement. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.04667' published_date: "2025-09-04" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - VC architecture: - GAN - hybrid relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - gan_adversarial_generation - hybrid_generation claims: - claim_id: darkstream_lookahead_content_latency_tradeoff role: supports claim: Limited lookahead in causal speech encoders can substantially improve linguistic-content preservation while retaining interactive latency. source: §V.A, §V.B, Tables I–II evidence: Wave+CL token accuracy rose from 53.16% at zero lookahead to 78.99% at 140 ms, with end-to-end latency increasing from 84.3 ms to 203 ms. confidence: high relevance: low - claim_id: darkstream_content_privacy_tension role: complicates claim: Improving linguistic clarity in streaming anonymization can weaken privacy because cleaner representations remain more discriminative for speaker recognition. source: §V.E, Table III evidence: Adding contextual layers lowered lazy-informed EER from 36.61% to 20.35% at zero lookahead, indicating greater speaker linkability. confidence: high relevance: low - claim_id: darkstream_quantization_privacy_quality_tradeoff role: complicates claim: Content-token quantization can improve anonymization toward chance-level verification while imposing substantial intelligibility and perceived-quality costs. source: §V.C, §V.D, §V.E, Tables IV–V evidence: A 256-centroid bottleneck raised lazy-informed EER to about 47%, while WER increased from 2.09% to 9.52% and MOS fell from 3.79 to 3.22. confidence: high relevance: high - claim_id: darkstream_streaming_matches_offline_privacy role: supports claim: Streaming voice anonymization can approach offline baseline privacy performance while maintaining real-time latency. source: §V.F, Table VI evidence: At 140 ms lookahead, DarkStream reported 22.68% semi-informed EER, comparable with offline VoicePrivacy baselines at 26.28% and 22.09%. confidence: medium relevance: low - claim_id: darkstream_mos_exposes_hidden_quantization_cost role: complicates claim: Human naturalness ratings can expose anonymization artifacts whose perceptual cost is understated by intelligibility metrics. source: §V.D, Table V evidence: Across a 20-listener test, quantization reduced MOS from 3.79 to 3.22 even though WER degradation alone did not capture the full perceived-quality loss. confidence: medium relevance: high limitations: - The MOS test involved 20 listeners, and perceptual quality was not compared against the offline VoicePrivacy baselines. - Evaluation is English-only and does not establish generalization to accented or code-switched speech. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.04685' published_date: "2025-09-04" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - codec architecture: - GAN relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - gan_adversarial_generation claims: - claim_id: content_adaptive_token_allocation_in_acoustic_tokenisers_achieves role: supports claim: Content-adaptive token allocation in acoustic tokenisers achieves better reconstruction quality than fixed-rate designs at equal or lower token budgets. source: §4.2, Table 1 evidence: VARSTok at 30.95 Hz achieves UTMOS 3.8949 on LibriTTS test-clean, surpassing the 40 Hz WavTokenizer (3.6107) while using 23% fewer tokens; the 36.81 Hz configuration (UTMOS 4.000) nearly matches the 75 Hz WavTokenizer (4.025) with fewer than half the tokens. confidence: high relevance: high - claim_id: dynamically_segmented_speech_tokens_carry_more_semantically_discriminative role: supports claim: Dynamically segmented speech tokens carry more semantically discriminative information than uniformly sampled tokens at the same average rate. source: §4.3, Table 2 evidence: All VARSTok configurations outperform the 40 Hz WavTokenizer on all four ARCH benchmark classification tasks (emotion, digit recognition, intent), despite operating at lower average frame rates. confidence: high relevance: low - claim_id: encoding_token_duration_implicitly_in_the_vq_codebook role: supports claim: Encoding token duration implicitly in the VQ codebook index eliminates the need for auxiliary duration predictors and preserves compatibility with autoregressive speech language models. source: §3.4, §4.4, Table 3 evidence: The implicit duration coding scheme maps each cluster's content index k and duration d to a single token ID D = (d-1)*K + k, enabling a standard cross-entropy autoregressive model to generate variable-rate token sequences without modification; MOS and WER improve over the fixed-rate baseline in zero-shot TTS. confidence: high relevance: high - claim_id: more_aggressive_temporal_compression_in_variable_rate_tokenisers role: complicates claim: More aggressive temporal compression in variable-rate tokenisers trades reconstruction quality for token efficiency beyond a practical compression threshold. source: §4.2, Table 1 evidence: Increasing S_max from 2 to 8 reduces the average frame rate from 46.5 Hz to 22.38 Hz but degrades UTMOS from 4.038 to 3.647 and PESQ from 2.069 to 1.453 on LibriTTS test-clean; the optimal configuration (tau=0.7, S_max=4 at 30.95 Hz) sits at the knee of this trade-off curve. confidence: high relevance: high - claim_id: inference_speed_in_autoregressive_speech_lms_depends_primarily role: refines claim: Inference speed in autoregressive speech LMs depends primarily on sequence length rather than vocabulary size, so variable-rate tokenisers with expanded vocabularies still accelerate decoding. source: §J, Table 5 evidence: VARSTok (tau=0.6) achieves RTF 0.487 versus 0.766 for the 40 Hz WavTokenizer baseline (36% speedup) despite expanding the token vocabulary from K to K*S_max = 16,384 entries, because shorter sequences reduce the dominant cost of attention computation over more function evaluations. confidence: high relevance: low limitations: - Evaluation is restricted to English (LibriTTS). The clustering algorithm relies on cosine similarity in a WavTokenizer embedding space trained on English read speech; whether the density-peak boundaries remain meaningful for other languages, spontaneous speech, or emotionally expressive styles is untested. - Speaker similarity under more aggressive compression (tau=0.6, 26.29 Hz) does show a statistically modest decline in objective SIM (0.880 vs 0.918 for the baseline), and while subjective SMOS remains comparable, the long-tail impact on voices far from the training distribution is unknown. Codebook collapse becomes severe for K above 4096 in the expanded index space, suggesting that very large vocabulary configurations require dedicated regularisation strategies not addressed here. The clustering algorithm is not differentiable, so joint end-to-end training with a downstream TTS model is not straightforward. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.05863' published_date: "2025-09-06" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - autoregressive-LM relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: - autoregressive_generation claims: - claim_id: preference_based_alignment_dpo_improves_intelligibility_in_multilingual role: supports claim: Preference-based alignment (DPO) improves intelligibility in multilingual autoregressive TTS systems beyond what supervised fine-tuning achieves. source: §4.3, §5.2, Table 2, Table 3 evidence: LatinX (DPO) reduces WER across nearly all 30 cross-lingual language pairs compared to the supervised fine-tuned baseline, and outperforms XTTSv2 in most pairs; Romanian-source conditions show particularly large gains (e.g., ro-to-es at 0.45% WER). confidence: high relevance: high - claim_id: automated_speaker_similarity_metrics_based_on_speaker_encoder role: complicates claim: Automated speaker similarity metrics based on speaker encoder embeddings do not reliably reflect human perceptual judgments of voice identity in zero-shot TTS. source: §5.2, §6, Table 4, Table 5 evidence: XTTSv2 achieves higher Sim-O scores than both LatinX models, yet human evaluators strongly prefer LatinX speaker similarity (SMOS 3.63/3.54 vs. 3.24); the paper explicitly flags this as a divergence between objective and subjective evaluation. confidence: high relevance: high - claim_id: dpo_alignment_in_tts_involves_a_trade_off role: complicates claim: 'DPO alignment in TTS involves a trade-off: optimizing for intelligibility and objective similarity can reduce naturalness MOS and, in some language conditions, perceptual similarity relative to the fine-tuned baseline.' source: §5.2, §6, Table 5, Table 6 evidence: LatinX (DPO) improves WER and Sim-E over the fine-tuned model but achieves lower average MOS (3.35 vs. 3.41) and lower SMOS in several cross-lingual conditions; the paper attributes this partly to the codec introducing artifacts that cap perceptual quality. confidence: high relevance: high - claim_id: lossy_neural_audio_codecs_set_a_perceptual_quality role: complicates claim: Lossy neural audio codecs set a perceptual quality ceiling in codec-based TTS that preference alignment cannot overcome, because the model learns to replicate codec artifacts introduced during reference encoding. source: §6 evidence: The paper notes that the VQ-VAE codec is lossy and the model learns to reproduce its artifacts, limiting the maximum perceptual quality achievable regardless of post-training alignment method. confidence: high relevance: high limitations: - The evaluation uses an internal test set of unseen speakers with no publicly named benchmark, and the human rating pool is predominantly English and Portuguese native speakers. Conclusions about multilingual naturalness and similarity, especially for French, Italian, and Romanian, should be treated with caution. - The DPO preference signal is constructed solely from WER and speaker similarity; no prosody, naturalness, or rhythm metric is incorporated, which likely explains the MOS regression relative to the fine-tuned baseline. The preference labeling is fully automated with no human verification of winner/loser assignments. The real-time factor of 4.85 makes the system unsuitable for real-time applications, and the authors note that non-autoregressive architectures are a necessary direction. The Romanian evaluation suffers from very small rater counts and predominantly non-native listeners, undermining the interpretation of the unusually high SMOS scores that exceed real audio. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.06074' published_date: "2025-09-07" entry_date: '2026-07-27' year: 2025 venue: EMNLP task: - TTS architecture: - transformer-enc-dec relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - transformer_encoder_decoder claims: - claim_id: word_level_semantic_and_prosodic_context_modeling_in role: supports claim: Word-level semantic and prosodic context modeling in dialogue history improves prosody expressiveness in conversational TTS over utterance-level context encoders. source: §3.5, Table 1; §3.6, Table 2 evidence: MFCIG-CSS achieves N-DMOS 3.980 and P-DMOS 3.899 on DailyTalk, outperforming seven baselines that operate at the utterance level by 0.122 and 0.104 respectively; ablation removing both graph modules causes the largest performance collapse across all metrics. confidence: high relevance: high - claim_id: graph_neural_networks_with_sequential_cross_turn_aggregation role: supports claim: Graph neural networks with sequential cross-turn aggregation can encode complementary semantic and prosodic interaction patterns from multimodal dialogue history for prosody-aware TTS. source: §3.6, Table 2 evidence: 'SIG and PIG are independently ablated: removing SIG reduces N-DMOS by 0.147 and P-DMOS by 0.106; removing PIG produces comparable drops. Both modules contribute distinctly and their combined use provides the strongest performance.' confidence: high relevance: high - claim_id: objective_energy_error_and_subjective_prosody_quality_metrics role: complicates claim: Objective energy error and subjective prosody quality metrics can rank conversational TTS systems differently. source: §3.5, Table 1 evidence: MFCIG-CSS achieves best N-DMOS (3.980) and P-DMOS (3.899) on DailyTalk but ranks second on MAE-E (0.314 vs. 0.310 for MSRGCN-CSS), indicating that signal-level energy accuracy does not fully predict human prosody quality judgments. confidence: high relevance: high - claim_id: prosody_modeling_gains_demonstrated_with_acoustic_feature_based role: complicates claim: Prosody modeling gains demonstrated with acoustic-feature-based TTS backbones may not transfer to codec-token-based or flow-based architectures. source: §5 Limitations evidence: MFCIG-CSS is validated only on a FastSpeech 2 backbone; the authors explicitly identify extension to VITS-based architectures and discrete token-based speech encoders as future work, acknowledging that the current validation scope limits generalizability claims. confidence: high relevance: low limitations: - MFCIG-CSS is evaluated on a single English dialogue dataset (DailyTalk, approximately 20 hours) with a FastSpeech 2 backbone, leaving generalization to other languages, longer conversations, noisy conditions, and modern autoregressive or codec-based TTS systems untested. The interaction graphs operate on frame-averaged acoustic features from Wav2Vec 2.0 and do not yet model fine-grained intra-word acoustic cues such as emotion, emphasis, or pauses. Extension to VITS-based architectures and discrete token-based speech encoders is the primary open direction identified by the authors. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.07038' published_date: "2025-09-08" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - singing architecture: - diffusion relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - diffusion_generation claims: - claim_id: explicit_conditioning_on_a_signal_derived_attribute_provides role: supports claim: Explicit conditioning on a signal-derived attribute provides substantially more controllability than implicit predictors trained on the same attribute as an auxiliary loss. source: §V.B, Table III evidence: Adding an energy predictor to the baseline reduces energy MAE from 0.33 to 0.30, while replacing the predictor with explicit phoneme-level energy input achieves 0.14 - a 57.6% reduction over baseline vs. 9% for the predictor. confidence: high relevance: low - claim_id: phoneme_level_aggregation_of_frame_level_features_enables role: supports claim: Phoneme-level aggregation of frame-level features enables user-friendly control in singing synthesis at acceptable precision cost. source: §III.B, §V.A, Table I, Table II evidence: Phoneme-level energy (L values, matching lyric and note sequence length) achieves energy MAE of 0.14 vs. 0.03 for frame-level (T values, roughly 1,000 per utterance), with higher MOS (3.78 vs. 3.57), demonstrating that coarser granularity trades minor precision loss for substantially reduced control burden. confidence: high relevance: high - claim_id: ground_truth_signal_features_can_substitute_for_manually role: supports claim: Ground-truth signal features can substitute for manually annotated expressive attributes as conditioning signals in singing voice synthesis. source: §I, §III.B evidence: Frame-level energy is extracted directly from ground-truth mel-spectrograms without human annotation, yet conditioning on it enables effective dynamics control, reducing annotation costs relative to technique-labelled datasets used in prior controllable SVS work. confidence: high relevance: high - claim_id: perceptual_quality_and_controllability_precision_do_not_align role: complicates claim: Perceptual quality and controllability precision do not align monotonically across conditioning granularities in SVS. source: §V.A, Table I, Table II evidence: The frame-level model achieves the best energy MAE (0.03) but lower MOS (3.57 ± 0.18) than the phoneme-level model (MOS 3.78 ± 0.19, energy MAE 0.14), suggesting that very tight energy matching may over-constrain generation in ways that reduce perceived naturalness. confidence: high relevance: high limitations: - MOS evaluation involved only 10 listeners, which the authors acknowledge constrains statistical reliability and generalizability. Results should be treated as directional rather than definitive. - The study uses only the Chinese subset of GTSinger and two speakers, leaving multilingual and speaker-diversity generalisability untested. The baseline DDPM architecture is explicitly noted to fall below state-of-the-art SVS performance, so the absolute quality figures do not reflect what the conditioning mechanism would achieve on a competitive backbone. Expressive attributes beyond dynamics (timbre, vibrato, advanced singing techniques) are not addressed. Phoneme-level energy may not capture fine-grained intra-phoneme temporal variation present in skilled vocal performance. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.07376' published_date: "2025-09-09" entry_date: '2026-07-27' year: 2025 venue: EMNLP task: - TTS architecture: - VAE relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - variational_latent_generation claims: - claim_id: hierarchical_facial_features_improve_voice_face_congruence role: supports claim: Hierarchical, locally-aware facial feature extraction substantially improves voice-face identity congruence compared to global visual representations in face-to-speech synthesis. source: §5.3, Table 5 evidence: Progressive Facial Granularity Aggregation (FGA), hierarchical patch decomposition with progressive convolutional aggregation, yields SECS of 79.96 (seen) vs. 66.25 when replaced with a single vanilla transformer over the full image, a gap of 13.71 points. confidence: high relevance: low - claim_id: crossmodal_attribute_supervision_improves_face_voice_alignment role: supports claim: Cross-modal multi-task supervision over demographic attributes improves face-voice alignment in end-to-end synthesis systems. source: §5.2, Table 4 evidence: Bilateral attribute enhancement (predicting gender and ethnicity from both the facial embedding and synthesized audio) produces the highest SECS configuration at 79.96 seen / 71.39 unseen; applying supervision to only one modality yields lower scores (77.76 / 70.27 for visual-only, 75.84 / 69.52 for audio-only). confidence: high relevance: low - claim_id: multiview_augmentation_critical_for_ftv_generalization role: supports claim: Multi-view face augmentation, pairing multiple facial frames per speaker across varying poses and lighting, is critical for robust speaker identity generalization in face-to-speech systems. source: §5.3, Table 5 evidence: Removing the multi-view augmentation strategy causes the largest single-component SECS drop in ablation studies (from 79.96 to 64.43, a decrease of 15.53 on seen speakers and 9.35 on unseen), exceeding the contribution of any individual architectural module. confidence: high relevance: low - claim_id: e2e_ftv_still_needs_crossmodal_semantic_supervision role: complicates claim: End-to-end face-to-voice synthesis without audio-space alignment targets still requires explicit cross-modal semantic supervision to maintain speaker identity coherence. source: §5.2, §5.3, Table 4, Table 5 evidence: Removing bilateral attribute supervision, even partially, consistently degrades SECS; visual attribute enhancement removal alone costs 5.23 SECS points. confidence: medium relevance: low - claim_id: ftv_systems_uneven_across_demographic_groups role: complicates claim: Face-to-voice synthesis systems trained on demographically skewed corpora may produce uneven speaker similarity across underrepresented groups. source: §Appendix B, Table 6, Table 7 evidence: The LRS3-TED training set is 74.66% male and 72.23% Caucasian; the proposed model achieves SECS of 75.97 for male and 75.03 for female speakers, while Middle Eastern (69.25) and African American (71.65) speakers show measurably lower scores despite the bilateral attribute supervision. confidence: high relevance: high limitations: - The evaluation uses 40 speakers and 21 annotators from LRS3-TED, which is heavily skewed toward Caucasian male speakers (74.66% male, 72.23% Caucasian), limiting statistical power for cross-demographic comparisons; comparison baselines do not include the most recent large-scale zero-shot TTS systems. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.09716' published_date: "2025-09-09" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: lalm_judges_approximate_human_style_evaluation role: supports claim: Large audio-language model judges can approximate human evaluation of speaking style adherence with near-human-level reliability. source: §3.3, Table 2 evidence: LALM-as-a-Judge (Gemini-2.5-pro) achieves Spearman correlations of 77.01% (English) and 73.03% (Chinese) with consensus human scores, comparable to inter-human agreement of 78.58% and 70.54%. confidence: high relevance: high - claim_id: opensource_slms_deficit_in_expressive_style_vs_commercial role: complicates claim: Current open-source spoken language models exhibit substantial deficits in controllable expressive speech generation compared to commercial systems. source: §3.2, Table 1 evidence: Open-source models (Step-Audio, Kimi-Audio, Qwen2.5-Omni, Baichuan-Audio) score 2.46-3.11 overall on VStyle, while commercial systems (GPT-4o, Doubao) reach 3.74-4.10, with the gap attributed to weak acoustic feature modeling and lower instruction-following robustness. confidence: high relevance: low - claim_id: style_control_uneven_across_languages role: complicates claim: Speaking style control does not transfer evenly across languages even within the same system, indicating language-specific rather than language-general style representations in current models. source: §3.2 evidence: 'Doubao and Kimi-Audio score substantially higher in Chinese than English (e.g., Doubao: 4.10 vs. 3.88), while the GPT-4o series shows the reverse, attributed to imbalanced training data distributions and differences in pronunciation habits.' confidence: medium relevance: low - claim_id: staged_hierarchical_evaluation_captures_finer_distinctions role: supports claim: Staged hierarchical evaluation frameworks capture finer distinctions in speech generation quality than flat or single-dimension scoring approaches. source: §2.3, Figure 2, §3.3 evidence: VStyle's LALM-as-a-Judge separates content correctness, style adherence, and naturalness into sequential assessment gates, enabling discrimination between systems that achieve content but fail on style versus those achieving both; this structure is validated against human evaluation at near-human-level correlation. confidence: high relevance: high limitations: - The instruction dataset is built from manually designed seeds expanded by LLMs, meaning its distribution reflects annotator preferences rather than natural user interaction needs; audio instructions are synthesized via commercial voice cloning, narrowing the style range in input prompts relative to naturally produced speech. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.08696' published_date: "2025-09-10" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - flow-matching relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - flow_matching_generation claims: - claim_id: smoothcache_zero_training_caching_accelerates_fm_tts role: supports claim: Zero-training transformer layer caching can accelerate diffusion-based TTS inference without retraining or architectural changes. source: §3.2, Table 1 evidence: Applying SmoothCache to F5-TTS with alpha=0.25 (caching approximately half of 32 NFE steps) reduces RTF from 0.46 to 0.26 while keeping WER, SIM-o, and UTMOS within measurement noise on LibriSpeech-PC test-clean. confidence: high relevance: high - claim_id: smoothcache_quality_benefit_depends_on_trajectory_length role: complicates claim: The inference-time quality benefit of layer caching in diffusion TTS depends on the length of the denoising trajectory. source: §4.1, Table 1 evidence: At 32 NFE, caching 50% of steps preserves quality; at 16 NFE, the same proportional caching degrades UTMOS from 3.90 to 3.79 and NISQA from 4.14 to 3.97 on LibriSpeech-PC test-clean, comparable to reducing NFE directly. confidence: high relevance: high - claim_id: smoothcache_independent_schedules_cause_artifacts role: complicates claim: Per-layer cache schedules derived independently for attention and feed-forward sublayers introduce inter-layer dependency artefacts in diffusion transformer TTS. source: §3.1 evidence: Applying separate Attn-only and FFN-only cache schedules to F5-TTS caused audible artefacts in informal listening tests; a unified schedule derived from the attention layer pattern resolved this issue. confidence: medium relevance: high - claim_id: smoothcache_caching_beats_step_reduction_at_high_nfe role: supports claim: Calibrated layer caching provides a marginally superior quality-compute trade-off compared to equivalent step reduction at high NFE settings. source: §4.1, Table 2 evidence: A 37-participant user study found 53% preference for 32 NFE with caching over a 24 NFE no-cache baseline at the same compute budget; the difference collapsed to 49% vs. 51% at 16 NFE. confidence: medium relevance: high limitations: - RTF measurements use a single GPU (NVIDIA T4) and a single flow-matching TTS model (F5-TTS); generalization to other DiT-based TTS systems or GPU architectures is untested. - The calibration set contains only ten utterances, which may not capture the full distribution of synthesis conditions. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2508.18240' published_date: "2025-09-15" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - SCA - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: semantic_nonverbal_capability_gap_in_s2s_llms role: supports claim: Current speech-to-speech language models exhibit a systematic capability gap between semantic understanding and non-verbal speech dimensions such as paralinguistic and ambient sound processing. source: §6.3.1, Table 2 evidence: On MTalk-Bench, all evaluated models outperform the human baseline on semantic rubric scores but underperform on paralinguistic and ambient sound dimensions, with no model exceeding 80/100 overall. confidence: high relevance: high - claim_id: multiturn_dialogue_verbosity_compensates_for_context_bottleneck role: complicates claim: Multi-turn spoken dialogue systems compensate for context-accumulation bottlenecks by increasing response verbosity rather than informational density. source: §6.3.2, Table 4 evidence: Turn-level rubric scores on MTalk-Bench show a non-linear dip from Turn 1 to Turn 2 with partial recovery at Turn 3, while content density declines linearly across turns even as token count increases. confidence: medium relevance: low - claim_id: modality_aware_design_outperforms_scale_in_multiturn_dialogue role: supports claim: Task-specific modality-aware design provides greater performance gains in multi-turn spoken dialogue than parameter scaling alone. source: §6.3.3, Table 2 evidence: Step-Audio-Chat (130B, with explicit text transcription of prior turns) achieves the highest overall Arena Elo (1049) among open-source models, while parameter count shows no reliable correlation with rubric or arena scores within the evaluated set. confidence: medium relevance: low - claim_id: llm_judge_unreliable_on_raw_audio_nonverbal_cues role: complicates claim: LLM-as-judge automated evaluation of speech quality is unreliable for non-verbal audio cues when models are given raw waveforms. source: §7.2.2 evidence: LLM evaluators achieve near-zero Spearman correlation with human evaluators on ambient sound and paralinguistic dimensions under raw-audio evaluation, while the same dimensions yield Spearman rho above 0.85 when non-verbal cues are provided as text annotations. confidence: high relevance: high - claim_id: arena_rubrics_agreement_requires_large_performance_gaps role: refines claim: Arena-style and rubrics-based evaluation methods yield consistent model rankings for speech systems, but only when performance gaps between models are sufficiently large. source: §7.2.1, §7.2.3, Table 2 evidence: Bootstrap rubric-exclusion analysis shows Spearman rho above 0.95 alignment between Arena and Rubrics rankings, but small ELO and rubric score differences correspond to statistically insignificant win-rate gaps. confidence: high relevance: low limitations: - The benchmark contains only 270 audio samples, and the paper's own analysis shows that Arena and Rubrics evaluations yield reliable distinctions only when performance gaps are large, a material constraint on discriminating between competitive models. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2412.16846' published_date: "2025-09-17" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - autoregressive-LM - VAE relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - autoregressive_generation - variational_latent_generation claims: - claim_id: distributional_training_objectives_for_continuous_ar_speech_modeling role: supports claim: Distributional training objectives for continuous AR speech modeling achieve higher intelligibility than regression-based alternatives. source: §TTS Evaluation, Table 2; §Ablation Study, Table 5 evidence: KALL-E with KL divergence loss achieves WER 1.94 / CER 0.96 on Seed-TTS test sets, below all discrete-token and regression-based baselines; ablation replacing Flow-VAE with Stable Audio VAE (near-zero KL weight, approaching a plain autoencoder) collapses CER from 2.79 to 40.09 at the same latent dimension. confidence: high relevance: low - claim_id: low_frame_rate_continuous_representations_reduce_autoregressive_tts role: supports claim: Low frame-rate continuous representations reduce autoregressive TTS inference compute by over an order of magnitude without sacrificing synthesis quality. source: §TTS Evaluation, Table 3, Table 4 evidence: KALL-E at 12.5 Hz requires 7,947 GFLOPs to synthesize 10 seconds vs. 122,170 for Llasa-1B at 50 Hz, while achieving higher MOS (4.17 vs. 3.92) and lower WER (1.94 vs. 3.6) on the same test set. confidence: high relevance: high - claim_id: objective_speaker_similarity_metrics_for_zero_shot_tts role: complicates claim: Objective speaker similarity metrics for zero-shot TTS are unreliable for cross-system comparisons when decoder architectures differ in their use of reference audio. source: §TTS Evaluation, Table 2, Table 3 evidence: Discrete-token systems (Seed-TTS SIM 0.796, FireRedTTS SIM 0.635) score differently on objective SPK-SIM than KALL-E (SIM 0.646/0.568), but KALL-E receives higher listener naturalness ratings; the authors attribute the gap to those systems conditioning the waveform decoder on the reference utterance at decode time, which inflates the metric independent of perceived speaker fidelity. confidence: high relevance: high - claim_id: test_time_adaptation_from_a_single_reference_utterance role: supports claim: Test-time adaptation from a single reference utterance improves speaker similarity in continuous-representation AR TTS without requiring full model retraining. source: §Test Time Training; §TTS Evaluation, Table 2 evidence: KALL-E (TTT) improves SPK-SIM from 0.568 to 0.611 on test-en using N=200 latent sequences sampled from the reference utterance's Flow-VAE distribution, with WER remaining stable at 1.90. confidence: high relevance: high - claim_id: increasing_vae_kl_regularization_weight_trades_reconstruction_fidelity role: complicates claim: Increasing VAE KL regularization weight trades reconstruction fidelity for a latent space structure that is more suitable for downstream generative modeling. source: §VAE Evaluation, Table 1; §Ablation Study, Table 5 evidence: Flow-VAE uses KL weight 32 and scores PESQ-WB 3.26 at 512-dim/12.5 Hz, below Stable Audio VAE (3.11) at the same frame rate with near-zero KL weight; however, Stable Audio VAE's latent space causes CER to collapse when used as the AR LM encoder, demonstrating that reconstruction quality and generation compatibility impose conflicting constraints on VAE training. confidence: high relevance: low limitations: - 'Objective speaker similarity remains below discrete-token systems that condition their decoders on the reference audio, suggesting the Flow-VAE''s information bottleneck trades some speaker detail for a more LM-friendly latent space. The TTT procedure assumes the transcript of the reference utterance is available, which may not hold in all deployment settings. Overfitting risk in TTT is real: CER rises after N=200 in ablation, limiting the effective adaptation set size. Evaluations are conducted solely on the Seed-TTS test sets; generalization to other benchmarks, out-of-distribution speakers, or noisy acoustic conditions is not assessed. Training data composition differs from the most directly comparable system (Llasa-1B), making it difficult to fully isolate architecture from data quality as the source of WER gains.' caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.13989' published_date: "2025-09-17" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: prosodic_control_does_not_scale_monotonically_with_system_capability role: complicates claim: Fine-grained prosodic control in instruction-guided TTS does not improve monotonically with overall system capability. source: §4.3, §4.4, Table 3 evidence: GPT-4o-mini-TTS led all five systems on adverbs-of-degree and emotion-intensity tasks yet achieved only 0.265 on word-level emphasis accuracy, barely above the 0.143 random baseline, and all systems showed near-zero F1 on child-voice generation regardless of system tier. confidence: high relevance: low - claim_id: human_perceptual_eval_reveals_gaps_automated_classifiers_miss role: supports claim: Human perceptual evaluation reveals ITTS alignment gaps that automated classifier-based metrics are likely to underestimate. source: §2.2, §3.3, §4.3 evidence: The E-VOC corpus collected over 60,000 ratings from 165+ raters on Prolific; large-scale human annotation found all five systems defaulted to adult-sounding voices regardless of explicit child or elderly instructions, a systematic bias that per-class classifiers tied to their training distribution would likely miss. confidence: high relevance: high - claim_id: systematic_adultvoice_bias_in_ITTS role: complicates claim: Current ITTS systems exhibit a systematic adult-voice bias that natural-language age instructions alone cannot overcome. source: §4.3, Table 3, Table 4 evidence: The GPT-4o confusion matrix shows utterances labeled Child, Teenager, and Elderly were predominantly perceived as Adult by human listeners, with the best child-voice F1 across all five systems only 0.113 and GPT-4o reaching only 0.074. confidence: high relevance: high - claim_id: commercial_ITTS_achieves_degree_scaled_control_open_source_cannot role: supports claim: Commercial ITTS systems can achieve perceptually ordered degree-scaled speech control that open-source alternatives fail to replicate at comparable levels. source: §4.1 evidence: GPT-4o-mini-TTS was the only system producing consistent objective acoustic progressions (LUFS, F0, words per second) and corresponding perceptual gradations under degree modifiers, while Parler-TTS variants, PromptTTS++, and UniAudio showed minimal, irregular, or reversed variation. confidence: high relevance: high limitations: - Annotation covers only native English speakers from the United States, and inter-rater agreement (Cohen's kappa) was low for the adverbs-of-degree (0.17) and emotion-intensity (0.226) tasks, indicating that even human raters find degree-scaled emotion perception genuinely ambiguous. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.14684' published_date: "2025-09-18" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - flow-matching relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - flow_matching_generation claims: - claim_id: daiental_crossattention_beats_concatenation role: supports claim: Injecting time-varying environment conditioning via cross-attention into flow-matching TTS backbones produces higher environment fidelity than direct concatenation of environment features. source: §4.2, Table 1 evidence: DAIEN-TTS with cross-attention achieves ESMOS 3.65 versus 3.49 for the concatenation-based ablated variant, with naturalness MOS also improving from 3.68 to 3.78. confidence: medium relevance: high - claim_id: daiental_separation_module_enables_environmental_prompts role: supports claim: A pretrained source separation module can enable zero-shot TTS to use environmental speech as a speaker prompt without degradation from background noise. source: §4.1, Table 1 evidence: DAIEN-TTS achieves WER 1.93% and MOS 3.84 under environmental speaker prompts, slightly outperforming F5-TTS with clean prompts (WER 2.30%, MOS 3.80) and substantially outperforming F5-TTS under the same environmental conditions (WER 2.87%, MOS 3.09). confidence: high relevance: high - claim_id: daiental_dual_cfg_independent_control role: supports claim: Dual classifier-free guidance enables independent control of speaker and environment attributes in a jointly trained flow-matching TTS model. source: §2.3.1, §4.2 evidence: DCFG with separate guidance strengths for the speech-text condition and the environment condition is the inference-time mechanism through which DAIEN-TTS independently controls timbre and acoustic background; ablating the cross-attention (which effectively collapses environment guidance) degrades environment fidelity without recovering speaker similarity. confidence: medium relevance: high - claim_id: daiental_no_baselines_for_timevarying_env role: complicates claim: The absence of established baseline systems for time-varying environment-aware TTS makes it difficult to assess absolute performance gains beyond ablation comparisons. source: §3.2 evidence: For the background environment synthesis condition, no existing systems match the task definition; all comparisons are between DAIEN-TTS and its own ablated variant, with no external reference point for environmental fidelity. confidence: medium relevance: low - claim_id: daiental_flowmatching_sensitive_to_prompt_environment role: complicates claim: Flow-matching TTS systems trained on clean speech are sensitive to acoustic environment in the speaker prompt, with degradation visible in both speaker similarity and naturalness. source: §4.1, Table 1 evidence: F5-TTS (trained on clean LibriTTS data) shows SIM-o dropping from 0.58 to 0.49 and MOS dropping from 3.80 to 3.09 when the speaker prompt contains background noise rather than clean speech. confidence: high relevance: high limitations: - Model size is not reported; the SES module introduces a separate pretrained component that must be trained and aligned with the TTS module, adding pipeline complexity; the SNR adaptation strategy assumes target SNR should match the environment prompt, which may not always align with creative intent. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.14946' published_date: "2025-09-18" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - autoregressive-LM - flow-matching relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: - autoregressive_generation - flow_matching_generation claims: - claim_id: automated_mixing_of_isolated_paralinguistic_audio_events_with role: supports claim: Automated mixing of isolated paralinguistic audio events with clean conversational speech, combined with voice conversion for timbre alignment, can produce effective fine-tuning data for paralinguistic TTS. source: §4.2.1, Table 3 evidence: CosyVoice2 SFT on SynParaSpeech achieves PMOS 3.31 versus 2.35 for NVS SFT and 1.88 baseline; F5-TTS SFT on SynParaSpeech similarly achieves PMOS 3.10 versus 1.49 for NVS SFT. confidence: high relevance: high - claim_id: dpo_with_naturally_paired_utterances_with_vs_without role: supports claim: DPO with naturally paired utterances (with vs. without paralinguistic events) can improve paralinguistic speech quality beyond SFT alone. source: §4.2.1, Table 3 evidence: CosyVoice2 DPO-Joint achieves PMOS 3.46 and QMOS 4.12 versus SFT-only PMOS 3.31 and QMOS 4.01; joint SFT+DPO outperforms staged DPO across all subjective dimensions. confidence: high relevance: high - claim_id: standard_automatic_speech_quality_metrics_are_unreliable_for role: complicates claim: Standard automatic speech quality metrics are unreliable for evaluating paralinguistic TTS systems. source: §4.2.1, Table 3 evidence: All models fine-tuned on paralinguistic data show degraded CER, SECS, and UTMOSv2 relative to baselines, despite achieving higher subjective NMOS and QMOS; laughter vocalizations raise CER because ASR transcribes them as "ha ha". confidence: high relevance: high - claim_id: few_shot_in_context_prompting_with_paralinguistic_examples role: supports claim: Few-shot in-context prompting with paralinguistic examples improves paralinguistic event detection in multimodal LLMs. source: §4.2.2, Table 4 evidence: SynParaSpeech prompt tuning raises accuracy and macro F1 for both Kimi Audio and Qwen 2.5 Omni over no-context baselines, with optimal performance at 5-shot context for both models. confidence: high relevance: low - claim_id: increasing_prompt_context_length_for_in_context_learning role: complicates claim: Increasing prompt context length for in-context learning in paralinguistic event detection yields diminishing returns and eventually degrades performance. source: §4.2.2, Table 4 evidence: Kimi Audio accuracy peaks at 5-shot (0.382) and declines at 7-shot (0.371 for Kimi Audio); Qwen 2.5 Omni also peaks at 5-shot (0.473) and declines at 7-shot (0.423), reflecting input overload beyond an optimal context window. confidence: high relevance: low limitations: - The corpus is Chinese-only and covers only 6 paralinguistic categories, limiting direct applicability to other languages or richer paralinguistic inventories. The paralinguistic events are inserted synthetically rather than captured in natural speech, which may introduce distribution mismatch relative to genuinely spontaneous vocalizations. - Standard objective metrics (CER, UTMOSv2, SECS) degrade after paralinguistic fine-tuning and do not reflect the true quality improvement observed in subjective evaluation. This leaves the field without reliable automatic metrics for paralinguistic speech, requiring human listener studies for valid comparisons. The paper does not evaluate zero-shot generalization to unseen paralinguistic categories or cross-lingual transfer. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2505.17093' published_date: "2025-09-19" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: [] relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: [] claims: - claim_id: llm_conversion_layer_reduces_expertise_burden_without_retraining role: supports claim: An LLM-based intermediate conversion layer can reduce the expertise burden on users of instruction-conditioned TTS without requiring model retraining. source: §4.3, Table 2 evidence: P2VA-C converts persona descriptions to structured prompts via GPT-4o-mini zero-shot prompting, improving human MOS from 3.09 to 3.42 and reducing WER from 22% to 17% relative to an unguided baseline on 1,000 LJSpeech-transcript persona pairs. confidence: high relevance: high - claim_id: llm_generated_voice_attributes_not_demographically_neutral role: complicates claim: Natural language voice attribute prompts generated by LLMs should not be assumed to be demographically neutral. source: §5.1, Table 3 evidence: When personas lack explicit gender cues, LLM-generated voice attributes assign male gender 64% of the time (vs. 10% in source descriptions), and 90% of accent assignments default to North American or British variants. confidence: high relevance: low - claim_id: structured_vs_freeform_tradeoff_extends_to_conversion_layers role: refines claim: The trade-off between structured and free-form style conditioning in instruction-conditioned TTS extends to intermediate conversion layers, not only the TTS interface itself. source: §4.3, Table 2 evidence: P2VA-C (structured, predefined attributes) achieves lower WER and higher human MOS than P2VA-O (free-form), while P2VA-O preserves expressivity for attributes outside the predefined vocabulary. confidence: medium relevance: high - claim_id: fairness_in_voice_generation_requires_addressing_llm_mediated_bias role: complicates claim: Fairness in voice generation systems requires addressing bias in LLM-mediated style inference, not only in training data or TTS model design. source: §5.1, §5.2, Tables 4-6 evidence: Stereotyped gender-tone and gender-pitch co-occurrences (e.g., male assigned Cognitive and Analytical tone 41%, female assigned Warm and Supportive 44% and high pitch 62%) emerge from the LLM conversion layer itself, independent of the downstream TTS model's behavior. confidence: high relevance: low limitations: - The user study does not report participant demographics, task instructions, or statistical significance, and the UTMOS improvement is acknowledged by the authors to be of uncertain statistical significance. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.15626' published_date: "2025-09-19" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - VAE relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - variational_latent_generation claims: - claim_id: libritts_vi_leakage_structural role: supports claim: Impression leakage in controllable TTS arises structurally when a single reference utterance is used for both speaker identity and style conditioning during training. source: §2, §5.2, Table 2 evidence: VIC-base achieves leakage delta 0.22 (significantly different from zero), confirming the reference audio biases synthesized VI toward the reference's inherent voice impression even when GRL and high-rate dropout are applied. confidence: high relevance: low - claim_id: libritts_vi_decoupled_utterances_reduce_leakage role: supports claim: Decoupled training using distinct utterances for speaker and style conditioning reduces style leakage without changing the underlying TTS architecture. source: §4.1, §5.2, Table 2 evidence: VIC-dis uses a separate utterance from the same speaker for reference conditioning during training, reducing leakage delta from 0.22 to 0.14 with a statistically significant improvement in RVI-MSE (0.61 to 0.51). confidence: high relevance: low - claim_id: libritts_vi_speakerfree_strongest_leakage_reduction role: supports claim: Eliminating the speaker reference and conditioning solely on a style vector achieves the strongest leakage reduction, at a moderate cost in speaker fidelity. source: §4.2, §5.2, Table 2 evidence: VIC-srf reduces the leakage delta to 0.05 (not significantly different from zero), while SECS drops to 0.72, remaining above the cross-speaker bound of 0.63 but below the same-speaker bound of 0.81. confidence: high relevance: low - claim_id: libritts_vi_llm_tts_poor_numerical_control role: complicates claim: LLM-based TTS systems conditioned on natural language prompts are inadequate for fine-grained numerical voice impression control. source: §5.1, §5.2, Table 2, Table 3 evidence: Qwen3-TTS (zero-shot) achieves VI-MSE of 0.82 versus 0.39 for the VITS-based baseline; fine-tuning worsens controllability further (VI-MSE 0.87), and text-VI entanglement is confirmed empirically by punctuation biasing predicted impressions. confidence: medium relevance: low - claim_id: libritts_vi_annotation_agreement_valid role: supports claim: Perceptual voice impression annotation can achieve inter-annotator agreement comparable to other subjective speech tasks, supporting the validity of VI corpora. source: §3, Table 1 evidence: Krippendorff's alpha averages 0.470 across 10 VI dimensions, exceeding reported agreement for speech emotion recognition (0.442) and singing voice preference (0.153). confidence: medium relevance: high limitations: - The LibriTTS-VI corpus contains only 130 manually annotated utterances; the remaining LibriTTS-R annotations are estimated by a trained VI estimator; several VI dimensions fall below the reliable inter-annotator agreement threshold; evaluation is confined to the clean LibriTTS-R audiobook domain; the subjective MOS evaluation uses only two speakers. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.15629' published_date: "2025-09-19" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - VC - singing - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: dynamic_singing_features_harder_to_convert_than_static_features role: supports claim: Dynamic temporal features in singing (vibrato, glissando, breathy voice) are substantially harder to convert than static spectral features, and current systems cannot model them reliably. source: §4.1 evidence: Style accuracy scores for vibrato (43.9%), glissando (42.6%), and breathy (37.3%) were markedly lower than those for falsetto (48.8%), mixed voice (48.3%), and pharyngeal (44.8%), with no system approaching human-level performance on the dynamic styles. confidence: high relevance: high - claim_id: joint_identity_style_conversion_degrades_naturalness role: supports claim: Jointly converting singer identity and singing style simultaneously degrades naturalness compared to identity-only conversion. source: §4.1 evidence: Top SVCC 2025 systems achieved approximately 3.7 MOS naturalness, below the approximately 3.9 MOS ground truth, whereas prior SVCC 2023 top systems reached human-level naturalness on the simpler identity-only task. confidence: medium relevance: high - claim_id: arlm_flowmatching_competitive_with_diffusion_for_svc role: supports claim: Autoregressive language model architectures with flow-matching decoders are competitive with specialized diffusion systems for singing voice conversion. source: §4.1 evidence: Systems based on Vevo1.5 (ARLM plus flow-matching decoder) achieved the highest singing style similarity scores in the challenge, outperforming pure diffusion-based systems using SeedVC and Serenade. confidence: medium relevance: low - claim_id: neural_mos_predictors_correlate_better_with_singing_naturalness role: supports claim: Neural MOS prediction models show stronger correlation with subjective naturalness scores for singing voice than traditional signal-based metrics. source: §4.2 evidence: SHEETSSQA and SingMOS achieved Spearman rank correlation coefficients above 0.6 with subjective naturalness scores, while speaker and singer embeddings were the strongest predictors of identity and style similarity (SRCC above 0.75). confidence: high relevance: high - claim_id: singer_identity_evaluation_confounded_by_reference_design role: complicates claim: 'Singer identity similarity evaluation in challenges is confounded by evaluation design: even ground truth recordings achieve only approximately 70% similarity scores with a single-sample reference.' source: §4.1 evidence: Ground truth samples reached only approximately 70% singer identity similarity in SVCC 2025 listening tests, prompting the challenge to present three reference samples per singer instead of one to reduce this floor effect. confidence: high relevance: high limitations: - Objective metrics show weak-to-moderate correlation with subjective scores across most dimensions, with only neural MOS predictors and embedding-based similarity reaching SRCC above 0.6, so objective-only comparisons for singing style conversion should be treated with caution until better proxy metrics are established. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.16010' published_date: "2025-09-19" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS - VC architecture: - hybrid relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - hybrid_generation claims: - claim_id: fedpisa_separate_adapters_improve_both role: supports claim: Disentangling speaker identity from expressive style into separate adapter modules improves both speaker similarity and style expressivity in personalised voice cloning. source: §3.3, Table 2 evidence: Ablations on GPT-SOVITS-V4 show removing the private ID-LoRA drops speaker similarity from 0.645 to 0.507, while removing the Style-LoRA drops style expressivity from 0.704 to 0.588; a single unified adapter cannot optimise both objectives without mutual degradation. confidence: high relevance: high - claim_id: fedpisa_collaborative_style_overcomes_scarcity role: supports claim: Collaborative style learning across clients in a federated setting can overcome per-client data scarcity and produce higher speaker similarity than purely local fine-tuning. source: §3.2, Table 2 evidence: Fed-PISA exceeds local fine-tuning in both speaker similarity (0.645 vs 0.529) and naturalness MOS (4.08 vs 3.36), demonstrating that sharing expressive style information across clients compensates for limited per-client data. confidence: medium relevance: high - claim_id: fedpisa_personalized_aggregation_beats_fedavg role: supports claim: Personalized aggregation weighted by stylistic similarity between clients outperforms uniform averaging for federated style learning. source: §3.3, Table 2 evidence: Under identical backbone and LoRA configuration, attention-based personalized aggregation achieves style expressivity 0.704 and speaker similarity 0.645 versus FedAvg's 0.476 and 0.523, confirming that style-averaged updates homogenise client models and harm expressivity. confidence: medium relevance: high - claim_id: fedpisa_stylisation_ceiling_without_identity_loss role: complicates claim: In speaker adaptation, stylisation steps beyond a small fraction of total training degrade speaker identity and peak naturalness, establishing a ceiling on achievable style transfer without identity loss. source: §3.4, Figure 2 evidence: Holding total steps at 100, speaker similarity decreases monotonically as stylisation steps increase, while naturalness peaks at 20% stylisation; excessive stylisation corrupts identity and degrades naturalness MOS. confidence: medium relevance: high - claim_id: fedpisa_legacy_backbone_comparison_confound role: complicates claim: Fair cross-framework comparison in federated TTS is obstructed when competing methods are architecturally coupled to legacy backbones and cannot be reimplemented on modern foundations. source: §3.1.2 evidence: FedSpeech and Federated Dynamic Transformer use FastSpeech2-XL and Transformer-TTS respectively and are incompatible with LoRA, making it impossible to isolate the contribution of the federation strategy from the backbone quality. confidence: medium relevance: low limitations: - Federated baselines run on architecturally distinct, weaker backbones, confounding attribution of gains to the federation strategy versus modern backbone quality; evaluation uses 60 speakers as 60 independent clients from four relatively small emotion datasets; naturalness is evaluated by only 22 in-lab researchers rather than crowd-sourced raters. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.16589' published_date: "2025-09-20" entry_date: '2026-07-27' year: 2025 venue: EMNLP task: - SCA - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: content_understanding_does_not_guarantee_direct_paralinguistic_recognition role: supports claim: Speech-LLMs that excel at content understanding do not reliably excel at direct paralinguistic recognition, revealing a systematic capability gap across model families. source: §4.2.1, Table 3 evidence: GPT-4o achieves 69.68 UAS on Content and 67.48 on Contextual Paralinguistic (long set) but only 30.34 on Direct Paralinguistic, while Qwen2-Audio reverses this pattern with 62.29 on Direct Paralinguistic-long but substantially lower Content and Contextual Paralinguistic scores. confidence: high relevance: low - claim_id: cascaded_pipelines_lose_paralinguistic_signal_vs_endtoend role: supports claim: Cascaded ASR-LLM pipelines lose critical paralinguistic signal compared to end-to-end speech-LLMs on tasks requiring direct or contextual paralinguistic reasoning. source: §4.2.1, Table 3 evidence: Cascade models score 33.3-44.18 on Direct Paralinguistic-long versus 60+ for end-to-end speech-LLMs, with the gap narrowing on direct-emotion-type questions where emotion cues are partly recoverable from transcripts. confidence: high relevance: low - claim_id: llm_judge_needs_multirun_averaging_for_paralinguistic_stability role: complicates claim: LLM-as-judge evaluation of paralinguistic reasoning requires multi-run averaging to achieve stable estimates, with single-run scores varying by up to 2.57 points per question type. source: §A.1.2, Table 5 evidence: LLaMA-3-70B judge scores showed standard deviations of 0.44-1.73 (long set) and 0.35-2.57 (short set) across five independent runs, with human-LLM correlation of 0.71 on a 27-pair subset of contextual paralinguistic questions. confidence: high relevance: high - claim_id: temperature_effects_on_speechllm_performance_are_task_and_length_dependent role: complicates claim: Inference temperature has task-dependent and input-length-dependent effects on speech-LLM performance, with no single optimal value across question types. source: §4.2.3, Figure 2 evidence: GPT-4o on contextual-gender-type questions peaks at mid-range temperatures, MERaLiON consistently favors lower values, and MERaLiON on contextual-speaker-count questions shows the largest gains with higher temperature on the short set. confidence: medium relevance: low - claim_id: shorter_inputs_generally_easier_except_for_direct_gender role: refines claim: Shorter speech inputs are generally easier for speech-LLMs across most question types, but direct gender inference does not follow this trend. source: §4.2.2 evidence: Models consistently score higher on the short set (10-30s) than the long set (30-60s) for Content, Contextual Paralinguistic, Contextual Emotion, Contextual Speaker-count, Direct Emotion, and Direct Speaker-count types, but direct-gender-type performance does not show the same length-dependent improvement. confidence: medium relevance: high limitations: - The benchmark is small in scale (under 500 samples per set), reducing statistical reliability for low-frequency question types, and QA pairs were generated by non-overlapping annotators without a cross-verification round, introducing potential individual annotation bias. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.20378' published_date: "2025-09-20" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - autoregressive-LM relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - autoregressive_generation claims: - claim_id: sub_sentence_emotion_conditioning_produces_more_accurate_emotional role: supports claim: Sub-sentence emotion conditioning produces more accurate emotional dynamics in TTS than sentence-level conditioning. source: §4.1, Table 1 evidence: Emo-FiLM outperforms CosyVoice2 by 9.1% on DTW and 12.7% on ESD DTW, with higher EMOS on both test sets, when comparing word-level emotion labels against global prompt conditioning. confidence: high relevance: high - claim_id: feature_wise_linear_modulation_is_an_effective_mechanism role: supports claim: Feature-wise Linear Modulation is an effective mechanism for injecting word-level conditioning signals into the text representations of autoregressive LLM-TTS systems. source: §4.2, Table 2 evidence: Ablation replacing the FiLM layer with simple addition increases FEDD DTW from 49.62 to 70.5, demonstrating that the non-linear affine modulation is essential beyond mere feature fusion. confidence: high relevance: low - claim_id: self_supervised_speech_emotion_models_provide_useful_word role: supports claim: Self-supervised speech emotion models provide useful word-aligned supervision for fine-grained emotional TTS when combined with forced alignment. source: §2.1, §4.2, Table 2 evidence: emotion2vec frame-level features aligned via MFA and pooled to word boundaries form the annotation backbone; removing word-level data tuning causes the most severe degradation (FEDD DTW 49.62 to 133.97). confidence: high relevance: high - claim_id: frame_averaged_emotion_similarity_metrics_are_insufficient_for role: complicates claim: Frame-averaged emotion similarity metrics are insufficient for evaluating intra-utterance emotional dynamics in TTS output. source: §3.4 evidence: The authors note that Emo SIM averages frame-level emotion vectors, potentially obscuring dynamic information, and introduce DTW as a complementary metric that more sensitively distinguishes the models' ability to track emotional transitions. confidence: high relevance: low - claim_id: explicit_emotion_classification_as_an_auxiliary_training_objective role: supports claim: Explicit emotion classification as an auxiliary training objective improves fine-grained emotion control in multi-task TTS training. source: §4.2, Table 2 evidence: Removing the emotion classification loss increases FEDD DTW from 49.62 to 73.96, a degradation comparable to removing the FiLM layer entirely. confidence: high relevance: low limitations: - The FEDD dataset is small (1,000 utterances, 5 speakers) and constructed partly via concatenation of emotionally distinct segments, which may not capture naturally occurring emotional transitions. The method is not evaluated for spontaneous or conversational speech, where emotion boundaries are less well-defined. The backbone (CosyVoice2) is frozen, so the approach inherits any limitations of that system in speaker diversity or prosodic range. Model size and speaker generalization are not reported. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.17021' published_date: "2025-09-21" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - autoregressive-LM relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - autoregressive_generation claims: - claim_id: exposure_bias_produces_measurable_quality_degradation_in_autoregressive role: supports claim: Exposure bias produces measurable quality degradation in autoregressive codec TTS, with the effect scaling with output sequence length. source: §3.4, Figure 3; §3.2, Table 1 evidence: Token prediction accuracy under free-running inference is consistently lower than under teacher forcing, with the gap widening over sequence position; WER improvement from hybrid training is larger on long-form LibriSpeech (6.23→4.21%) than on short Seed-TTS utterances (4.83→4.64%). confidence: high relevance: low - claim_id: hybrid_training_that_mixes_teacher_forced_and_self role: supports claim: Hybrid training that mixes teacher-forced and self-generated tokens reduces WER and improves speaker similarity in autoregressive LM-based TTS relative to standard teacher-forcing fine-tuning. source: §3.2, Table 1; §3.3, Figure 2 evidence: Prompt-guided hybrid fine-tuning of CosyVoice2 on LibriSpeech reduced WER from 6.23% to 4.21% and raised speaker similarity from 0.74 to 0.78 on LibriSpeech test-clean (Table 1); human MOS scores approach ground-truth quality on 30-sample evaluation. confidence: high relevance: high - claim_id: the_benefits_of_exposure_bias_mitigation_in_autoregressive role: complicates claim: The benefits of exposure-bias mitigation in autoregressive TTS diminish for short utterances where prediction errors have less opportunity to accumulate. source: §3.2, Table 1 evidence: On Seed-TTS utterances under 10 seconds, WER improvement is smaller (4.83→4.64) compared to LibriSpeech, where longer sequences amplify the compounding effect of distributional mismatch. confidence: high relevance: low - claim_id: eos_misprediction_rate_serves_as_a_reliable_proxy role: supports claim: EOS misprediction rate serves as a reliable proxy for training-time exposure bias severity, enabling adaptive control of self-conditioning intensity. source: §3.5; §3.6, Table 2 evidence: EOS-guided adaptive scheduling tracks premature termination events across iterations; ablation shows removing this component raises WER from 4.21% to 4.98% and drops speaker similarity from 0.80 to 0.72 on LibriSpeech test-clean. confidence: high relevance: high limitations: - The evaluation is limited to LibriSpeech and Seed-TTS with a single speaker conditioning setup; generalisation to multilingual or highly expressive TTS domains is untested. The method is applied only as fine-tuning on top of existing CosyVoice models, leaving open whether hybrid training from scratch would yield similar or greater benefits. The MOS evaluation uses only 30 samples, which provides limited statistical power for assessing naturalness improvements. The 1.5x training overhead, while modest, is an additional cost relative to standard fine-tuning. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.17143' published_date: "2025-09-21" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - VC architecture: - autoregressive-LM relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - autoregressive_generation claims: - claim_id: in_zero_shot_voice_conversion_temporally_coarser_syllabic role: supports claim: In zero-shot voice conversion, temporally coarser syllabic representations reduce pitch leakage from linguistic features more effectively than standard frame-aligned SSL features, enabling cleaner prosody control at the cost of intelligibility. source: §2.1, §4, Table 2 evidence: MaskVCT-Spk using SylBoost syllabic tokens achieves the lowest FPC (0.167) among all tested systems, indicating near-complete pitch independence from the source, while MaskVCT-All with continuous features retains more pitch correlation (FPC 0.417). confidence: high relevance: low - claim_id: multiple_classifier_free_guidance_weights_applied_to_distinct role: supports claim: Multiple classifier-free guidance weights applied to distinct conditioning factors in a single masked generative model enable user-configurable inference-time trade-offs between intelligibility, pitch fidelity, and speaker similarity. source: §2.5, §3.4, Table 2 evidence: MaskVCT defines triple CFG weights (w_all, w_spk, w_ling) over speaker, pitch, and linguistic conditions within one trained model; sweeping these weights continuously interpolates between MaskVCT-All (WER 4.68%, S-SIM 0.865) and MaskVCT-Spk (WER 6.47%, S-SIM 0.895). confidence: high relevance: high - claim_id: syllabic_speech_representations_that_suppress_pitch_leakage_in role: complicates claim: Syllabic speech representations that suppress pitch leakage in voice conversion also degrade content intelligibility through syllable misreadings caused by coarse temporal quantisation. source: §4, §5, Table 2 evidence: MaskVCT-Spk achieves the highest speaker similarity (S-SIM 0.895) but the highest WER (6.47%) among systems tested, substantially above FACodec (3.55%) and FreeVC (3.96%); the conclusion section attributes misreadings to K-means syllable mapping errors in SylBoost. confidence: high relevance: high - claim_id: masked_non_autoregressive_codec_models_can_match_or role: supports claim: Masked non-autoregressive codec models can match or exceed autoregressive and diffusion-based baselines on speaker similarity and quality in zero-shot VC while operating with fewer discrete tokens per utterance. source: §3.4, §4, Table 2 evidence: MaskVCT-Spk (2048 tokens) achieves higher S-SIM (0.895) and SS-MOS (3.69) than MaskGCT-S2A (8192 tokens, S-SIM 0.863, SS-MOS 3.02) and competitive UTMOS (3.17 vs. 3.24). confidence: high relevance: high limitations: - Syllabic tokens introduce misreadings that WER alone cannot fully diagnose; the authors acknowledge the K-means quantiser cannot recover from incorrect syllable boundary assignments, and propose future work to address this with a trainable VQ module. - The model is English-only. Accent conversion experiments are limited to L2-ARCTIC and test only two conversion directions; how well the syllabic pitch-stripping generalises to tonal languages (where pitch is phonemic) is unexplored. The 511-pair test set is relatively small for statistical confidence, especially given the reported confidence intervals overlap for several key comparisons. No code or trained checkpoint is publicly released, limiting reproducibility. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.17988' published_date: "2025-09-22" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - flow-matching relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - flow_matching_generation claims: - claim_id: in_the_wild_pipelines_produce_usable_tts_corpora_for_low_resource_languages role: supports claim: In-the-wild recordings can be converted into usable TTS training corpora at scale for languages lacking studio-recorded resources, provided the processing pipeline applies dedicated noise, diarization, VAD, and perceptual-quality filtering. source: §3.1, §3.2, Table 1 evidence: An adapted Emilia-style pipeline (denoising, Pyannote diarization, Silero VAD, DNSMOS P.835 OVRL >= 3.0 filtering) applied to parliamentary recordings yields 900 hours of Finnish and 5090 hours of Swedish TTS-ready speech, where previously only 60 hours (Finnish) or zero hours (Swedish) of public data existed. confidence: high relevance: high - claim_id: cross_validating_two_asr_models_filters_noisy_transcripts role: supports claim: Cross-validating transcripts from two independently trained ASR models is a practical way to filter noisy automatic transcripts for corpus construction when a single ASR system is unreliable for the target language. source: §3.1 evidence: Because faster-whisper-large-v3 underperformed on Finnish, the pipeline retained only segments where Whisper and a Finnish-specific Wav2Vec2-large model's transcripts agreed within 5% WER, using Whisper's transcript as the final label. confidence: medium relevance: low - claim_id: explicit_alignment_improves_intelligibility_at_naturalness_cost role: complicates claim: Explicit duration/alignment modeling in non-autoregressive TTS improves intelligibility and reduces hallucination relative to implicit alignment learned via diffusion transformers, but this comes at some cost to perceived naturalness. source: §5, Table 2 evidence: Matcha-TTS (explicit MAS alignment) achieves substantially lower CER (2.55% Finnish, 4.66% Swedish) and far fewer hallucinated utterances than F5-TTS-Base (6.72% Finnish, 13.64% Swedish CER; up to 18/500 hallucinations), yet F5-TTS-Base scores higher (better) CMOS in both languages, showing CER and CMOS can rank systems oppositely. confidence: high relevance: high - claim_id: speaker_similarity_metric_ranking_inconsistent_across_languages role: complicates claim: The relative speaker-similarity performance of prompt-based zero-shot conditioning versus fixed speaker-embedding conditioning is not consistent across languages or training corpora, and results from one language should not be assumed to transfer to another. source: §5, Table 2 evidence: On the same two model architectures, fixed speaker-embedding conditioning (Matcha-TTS) outperforms prompt-based conditioning (F5-TTS-Base) on Finnish SIM (0.566 vs 0.538) but underperforms it on Swedish SIM (0.442 vs 0.53). confidence: high relevance: low limitations: - The corpus is drawn entirely from formal parliamentary proceedings, a narrow speech register, and subjective evaluation used only 20-22 raters per language, a modest panel size for CMOS/SMOS testing. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.18470' published_date: "2025-09-22" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - diffusion - flow-matching relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - diffusion_generation - flow_matching_generation claims: - claim_id: graddt_discrete_time_matches_continuous_baseline role: supports claim: A fully discrete-time training and inference formulation can match the subjective and objective speech quality of a continuous-time diffusion baseline while eliminating the continuous/discrete mismatch inherent to score-based samplers. source: §4.1, Table 1; §4.2, Table 3 evidence: Grad-TTS-DT (discretized additive-Gaussian noising, clean-data prediction) achieves MCD 5.53, log F0 error 0.31, UTMOSv2 3.95 versus the continuous Grad-TTS baseline's 5.71/0.33/4.03 at 10 inference steps, and a higher subjective MOS (4.07 vs. 4.02). confidence: high relevance: high - claim_id: graddt_multiplicative_noise_viable_alternative role: supports claim: Multiplicative Gaussian noise is a viable alternative to additive Gaussian noise as a diffusion-like noising process for speech, despite its non-isotropic covariance structure. source: §4.1, Table 1, Table 2 evidence: RFMG (multiplicative noise) achieves MCD, log F0, and UTMOSv2 close to RFAG (additive noise) across matched sigma values at both 10-step and 5-step inference. confidence: high relevance: high - claim_id: graddt_deterministic_noising_insufficient_naturalness role: complicates claim: Fully deterministic noising processes are insufficient on their own for naturalness in discrete-time diffusion-like speech synthesis; some stochasticity in the noising process appears necessary. source: §4.1, Table 1, Table 2; §4.2, Table 3 evidence: The fully deterministic Blurring system records the lowest UTMOSv2 of all compared systems (3.71 at 10 steps, 3.66 at 5 steps) and ties for the lowest subjective MOS (3.86). confidence: medium relevance: high - claim_id: graddt_traintest_consistency_determines_robustness role: refines claim: Consistency between training and inference conditions, not just the number of inference steps, determines how well a diffusion-like TTS model generalizes to fewer sampling steps. source: §4.1, Table 1, Table 2 evidence: All discrete-time systems retain near-identical MCD and log F0 error when inference steps are reduced from 10 to 5, which the authors attribute to the discrete-time formulation exposing the model to identical conditions at training and inference. confidence: medium relevance: low limitations: - All systems are trained/evaluated on a single English, single-speaker corpus (LJSpeech) with a text encoder inherited unmodified from a pretrained Grad-TTS checkpoint the authors describe as suboptimal. - The paper does not compare against more recent flow-matching TTS systems (e.g. rectified-flow or OT-CFM models), only against the original score-based Grad-TTS, leaving unclear how the discrete-time variants compare to the current flow-matching state of the art. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.18531' published_date: "2025-09-23" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - autoregressive-LM relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: - autoregressive_generation claims: - claim_id: reinforcement_learning_with_automatically_verifiable_rewards_can_optimize role: complicates claim: Reinforcement learning with automatically verifiable rewards can optimize the measured objective while degrading unmeasured perceptual qualities such as prosodic naturalness. source: §4.3, Table 1, Figure 2 evidence: GRPO trained on a harmonic-mean reward of CER and token likelihood reduced CER to the lowest value among all tested systems (2.20%) but produced the lowest human-preference ELO (753.7), with log-F0 pitch distributions showing collapse toward monotone speech relative to the baseline. confidence: high relevance: high - claim_id: adding_an_auxiliary_reward_term_intended_to_fix role: complicates claim: Adding an auxiliary reward term intended to fix a known RL side effect can destabilize training instead of correcting the targeted attribute. source: §4.4 evidence: Extending the GRPO reward with a speaker-similarity utility increased similarity scores but inflated CER to 42.63% and produced degenerate non-terminating outputs (missing end-of-sequence tokens), consistent with reward hacking rather than genuine prosody improvement. confidence: high relevance: low - claim_id: small_scale_iterative_human_preference_optimization_can_recover role: supports claim: Small-scale iterative human preference optimization can recover perceptual speech qualities lost to metric-driven RL training without requiring a large static preference dataset. source: §4.5, Table 1 evidence: Using approximately 200 human preference pairs per round with a moving DPO reference, three iterative rounds raised ELO from 753.7 (GRPO) to a peak of 1190.1 at round 2, exceeding three commercial TTS baselines in human preference, while CER recovered to 3.30-3.60% by rounds 2-3. confidence: high relevance: high - claim_id: the_benefit_of_successive_rounds_of_iterative_preference role: refines claim: The benefit of successive rounds of iterative preference optimization is not monotonic and can saturate or reverse as the policy converges toward its own reference. source: §4.5 evidence: ELO peaked at round 2 (1190.1) and declined at round 3 (1064.2); the authors attribute this to a narrowing policy-reference gap reducing the informativeness of newly collected preference pairs in later rounds. confidence: high relevance: high limitations: - Beyond the evaluation scope, the paper does not report an ablation over the DPO sharpness parameter (beta) or over preference-pair budget per round, so it is unclear how sensitive the round-2 peak is to these choices. The mechanism behind the round-3 ELO decline is offered as a hypothesis (narrowing policy-reference gap) rather than directly verified. The GRPO speaker-similarity failure is also reported as a single run rather than across multiple reward-weighting configurations, leaving open whether a different weighting could avoid the observed instability. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.18823' published_date: "2025-09-23" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - evaluation - codec architecture: - GAN relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: - gan_adversarial_generation claims: - claim_id: fad_more_reliable_than_mmd_for_perceptual_quality role: supports claim: Fréchet-distance metrics computed on audio embeddings correlate more reliably with human perceptual quality judgments than kernel-based maximum mean discrepancy metrics. source: §4.1, Table 1 evidence: Across every embedding space tested (EnCodec, DAC, DACe, CLAP, CLAP LAION, OpenL3), FAD achieved equal or higher Pearson and Spearman correlation with MUSHRA scores than MMD on the same embeddings, e.g. DACe FAD Rp = 0.70 vs. DACe MMD Rp = 0.65. confidence: high relevance: high - claim_id: higher_fidelity_codec_embeddings_track_quality_better role: supports claim: Higher-fidelity neural audio codec embeddings track human-perceived audio quality more accurately than lower-fidelity codec embeddings. source: §4.1, Table 1; §3, Figs. 1-2 evidence: 'Correlation with MUSHRA scores rose monotonically with codec quality: EnCodec (Rp = 0.38) < DAC 16 kb/s (Rp = 0.68) < DACe (Rp = 0.70) under FAD, mirrored by a separate subjective MUSHRA test confirming DACe > DAC > EnCodec.' confidence: high relevance: high - claim_id: codec_embeddings_lag_general_purpose_embeddings_for_quality_prediction role: complicates claim: Codec-derived embeddings, even from high-fidelity codecs, are not automatically competitive with embeddings trained for perceptual or semantic similarity when used for generative audio quality prediction. source: §4.2 evidence: General-purpose CLAP LAION Music and OpenL3-128M embeddings outperformed the best neural codec embedding (DACe) in both Pearson and Spearman correlation with MUSHRA scores, attributed to codecs' reconstruction-based training objective and roughly 10x less training data than the contrastively/self-supervised-trained alternatives. confidence: high relevance: high - claim_id: mmd_sensitive_to_bandwidth_hyperparameter role: complicates claim: Kernel-based statistical distance metrics for audio evaluation are sensitive to bandwidth hyperparameter choice in a way that moment-based distances are not. source: §4.1 evidence: Fixed MMD kernel bandwidths (sigma = 1, 10, 100, 1000, 10000) tested on DACe embeddings peaked in correlation near sigma = 100 but never reached the correlation achieved with the median-distance heuristic bandwidth, with performance degrading more sharply for undersized than oversized bandwidths. confidence: medium relevance: low limitations: - Validation is performed in a signal-to-signal, reference-based setting (coded audio vs. its own uncoded reference under classical lossy coding artifacts), not on true generative-model outputs, so the reported correlations characterize embedding sensitivity to coding artifacts rather than to generative-model artifact types. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.18928' published_date: "2025-09-23" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - autoregressive-LM - diffusion - hybrid relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: - autoregressive_generation - diffusion_generation - hybrid_generation claims: - claim_id: ardm_dpo_extension_to_diffusion role: supports claim: Direct Preference Optimization can be extended beyond discrete-token generative models to continuous-token autoregressive diffusion models for speech, by marginalizing the DPO objective over intermediate diffusion states of the autoregressive sampling trajectory. source: §3, §4.2, Table 1 evidence: The derived ARDM-DPO objective reduces to a log-sigmoid of the difference in denoising losses between the fine-tuned and reference model on winning vs. losing trajectories; applied to a 0.4B DiTAR model it nearly doubles F0 variance (14.2 to 29.2) while WER improves (5.17 to 3.73) after 200 training steps. confidence: high relevance: high - claim_id: ardm_dpo_automatic_proxy_preference_data role: supports claim: Automatic proxy signals, rather than human-labeled preference pairs, can be used to construct DPO training data that improves a targeted, measurable aspect of speech generation quality. source: §4.3, Table 2 evidence: Task B builds 430k preference pairs by ranking 16 candidate outputs per prompt using a separately trained CTC model's negative log-likelihood as an intelligibility proxy; this reduces CER by 25% (8.37 to 6.32) on the Seed-TTS-Eval hard test set while preserving speaker similarity and naturalness in subjective tests. confidence: high relevance: high - claim_id: ardm_dpo_training_instability role: complicates claim: DPO fine-tuning of autoregressive diffusion speech models is not guaranteed to be stable and can require early stopping to prevent degradation of speaker similarity and naturalness. source: §4.2, Fig. 3, §5 evidence: Across the beta in {200, 400, 800} grid search for Task A, speaker similarity decreases monotonically throughout training regardless of beta, and the diffusion loss for both preferred and dispreferred samples increases during training instead of diverging as the objective predicts. confidence: high relevance: high - claim_id: ardm_dpo_beats_bestofk_and_raft role: refines claim: Preference-optimizing a generative TTS model via a single fine-tuning run can match or exceed the quality gains obtained by inference-time best-of-K sampling or iterative rejection-sampling fine-tuning, without their added inference-time or multi-round training cost. source: §4.2, Table 1 evidence: On Task A, ARDM-DPO (200 steps, beta=200) reaches F0V 29.2, higher than best-of-64 sampling (26.6, no training) and three iterations of RAFT rejection-sampling SFT (20.1), while requiring a single training run and no extra inference-time sampling. confidence: high relevance: high limitations: - All experiments fine-tune a single proprietary 0.4B DiTAR checkpoint pretrained on an internal 280,000-hour corpus, with no code or model release; the reported gains cannot be independently reproduced or verified on a public base model. - Training instability on Task A (monotonically decreasing speaker similarity across all tested beta values) is acknowledged as unresolved and requires manual early stopping rather than a principled stopping criterion. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.19231' published_date: "2025-09-23" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS - VC - evaluation architecture: - diffusion - GAN relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - gan_adversarial_generation - diffusion_generation claims: - claim_id: chiressd_style_disentanglement_suppresses_mispronunciation role: supports claim: Disentangling acoustic (pronunciation-linked) style from prosodic and speaker-identity style in a TTS-based reconstruction system allows mispronunciation patterns to be suppressed while speaker identity is preserved. source: §5.1, Table 1 evidence: ChiReSSD, which selectively fine-tunes acoustic/prosodic style encoders and the pitch extractor while reusing StyleTTS2's disentangled style representations, achieves higher speaker similarity (0.62) and lower pitch deviation (19.98%) than a one-shot baseline that transfers style holistically without domain adaptation (0.52 similarity, 22.58% F0 difference). confidence: medium relevance: high - claim_id: chiressd_asr_metrics_conflict_with_identity role: complicates claim: Optimizing a speech reconstruction system for automatic-speech-recognition-based intelligibility metrics can conflict with optimizing for speaker-identity preservation. source: §5.2, Table 2 evidence: A single-speaker adult-voice TTS baseline with no style transfer obtains the lowest WER/CER on STAR (0.34/0.11) precisely because it discards the child speaker's identity, while the identity-preserving ChiReSSD model scores worse on the same ASR-based metrics (WER 0.49, CER 0.29) despite being judged clinically more accurate. confidence: medium relevance: low - claim_id: chiressd_automatic_pcc_approximates_clinical role: supports claim: An automatic phone-recognition-based estimate of consonant production accuracy can approximate expert clinical judgments of speech intelligibility well enough to reduce, though not eliminate, the need for manual phonetic transcription. source: §5.3 evidence: Automatic PCC estimates derived from a universal phone recognizer and Levenshtein distance correlate with a certified speech-language therapist's manual PCC annotations at Pearson correlation 0.63 across 21 sample pairs. confidence: medium relevance: high - claim_id: chiressd_generalizes_across_disorder_populations role: supports claim: A style-based TTS reconstruction approach adapted for one clinical population of disordered speech can generalize to a structurally different disorder and age group without population-specific retraining. source: §5.4, Table 3 evidence: ChiReSSD, trained only on pediatric SSD data, reduces CER to below 0.03 and WER to as low as 0.06 on the TORGO adult dysarthria dataset across severity levels, while keeping speaker similarity at 0.74-0.77. confidence: medium relevance: high limitations: - Clinical validation of the automatic PCC metric rests on only 21 annotated sample pairs from a single speech-language therapist, with no inter-rater reliability reported; the STAR evaluation set is small (234 utterances from six children); the guidance parameters (alpha, beta) were tuned rather than learned. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.19812' published_date: "2025-09-24" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - GAN relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - gan_adversarial_generation claims: - claim_id: progressive_distillation_compresses_watermark_encoder_matching_robustness role: supports claim: Progressive, mixing-based knowledge distillation can compress a deep audio watermarking encoder to a fraction of its computational cost while matching the teacher model's robustness and imperceptibility. source: §IV.D-E, Tables I-II evidence: Distilling from an 8-block/32-channel INN teacher to a 2-block/16-channel student lowers GFLOPS by 93.6% (36.0 to 2.3) yet the student's average BER (0.51%) is lower than the teacher's (0.56%) and its PESQ (4.30) is within 0.03 of the teacher's (4.33). confidence: high relevance: low - claim_id: output_only_distillation_transfers_watermark_robustness_poorly role: complicates claim: Distilling a compressed watermarking encoder by supervising it only through the teacher's decoder output, without intermediate mixing of encoder representations, transfers robustness poorly. source: §IV.E, Table I evidence: A direct knowledge distillation (DKD) student trained under identical settings to PKDMark shows more than triple the bit error rate (1.94% vs 0.51%) and lower perceptual quality (PESQ 4.08 vs 4.30). confidence: high relevance: high - claim_id: complex_stft_domain_embedding_improves_watermark_robustness role: refines claim: Embedding watermark information using both phase and amplitude in the complex STFT domain, rather than amplitude alone, improves the robustness of deep audio watermarking against signal distortions. source: §IV.E, Table I evidence: Removing the complex-valued message feature map from the teacher model (amplitude-only embedding) raises average BER from 0.56% to 1.77% and drops SNR from 45.75dB to 42.25dB. confidence: high relevance: low - claim_id: basic_distortion_robustness_does_not_imply_advanced_attack_robustness role: complicates claim: Efficient watermark detectors that achieve near-perfect bit accuracy under basic signal distortions can still fail under advanced temporal-domain attacks such as time-stretching, pitch-shifting, or reverberation. source: §IV.C, Table I evidence: AudioSeal attains 0.00% BER on several basic distortions but its BER rises to 47-59% under slow-speed, fast-speed, and pitch-shift attacks, driving its average to 18.66% versus PKDMark's 0.51% under the same 14-distortion suite. confidence: high relevance: low limitations: - The paper explicitly leaves neural codec compression and neural denoising attacks untested; the evaluation compares against only two open-source baselines (WavMark, AudioSeal) and not several more recent efficient watermarking systems it cites. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.19883' published_date: "2025-09-24" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - singing architecture: - hybrid relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - hybrid_generation claims: - claim_id: comelsinger_prompt_leakage_language_independent role: supports claim: Prompt-based conditioning in masked generative or codec-language speech synthesis models causes measurable leakage of prosodic attributes from the acoustic prompt into the synthesized output, independent of the target language. source: §V.A, Table I evidence: Paired-prompt outputs show consistently lower pitch/energy/jitter differences than unpaired-prompt outputs from the same speaker on both LibriTTS (English) and AISHELL-3 (Mandarin) when synthesizing with MaskGCT. confidence: high relevance: low - claim_id: comelsinger_contrastive_reduces_leakage role: supports claim: Explicit contrastive regularization between the acoustic prompt and an external control signal (e.g., melody/pitch) reduces attribute leakage and improves controllability in prompt-based zero-shot synthesis. source: §V.D, Table V evidence: Removing the coarse-to-fine (sequence + frame level) contrastive loss increases F0-RMSE from 0.042 to 0.08 and lowers SingMOS from 4.32 to 4.12 on the seen-singer test set, with sequence-level and frame-level components independently ablated to show complementary effects on speaker-identity and pitch-detail metrics respectively. confidence: high relevance: high - claim_id: comelsinger_inloop_transcription_supervision role: supports claim: Integrating auxiliary transcription-derived frame-level supervision directly into a synthesis model's training loop, rather than using it only as an offline data-cleaning step, improves fine-grained attribute alignment. source: §V.D, Table V evidence: Removing the in-loop SVT auxiliary loss produces the largest single-component degradation in the ablation, raising F0-RMSE from 0.042 to 0.194 and lowering SingMOS from 4.32 to 3.95. confidence: high relevance: high - claim_id: comelsinger_lora_matches_full_finetuning role: refines claim: Parameter-efficient fine-tuning can match or exceed full fine-tuning when adapting a large pretrained codec-based speech model to a lower-resource downstream domain, provided the low-rank capacity is placed appropriately. source: §V.D, Table VII evidence: LoRA fine-tuning of the S2A diffusion estimator (6.51% trainable parameters) achieves lower F0-RMSE (0.053) and higher speaker similarity (0.92) than fully fine-tuning the same backbone (100% trainable, F0-RMSE 0.099, SECS 0.859). confidence: medium relevance: high limitations: - The zero-shot evaluation set (OpenSinger) lacks native music-score annotations, so pitch/duration sequences are paired from a different corpus (M4Singer); evaluation is confined to Mandarin singing corpora; overall S2A/T2S model size is not reported; the subjective evaluation panel is modest (20 musically trained raters). caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.19928' published_date: "2025-09-24" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: token_edit_distance_correlates_better_with_prosody_perception_than_acoustic_proxies role: supports claim: Acoustic proxy metrics for prosodic variation correlate weakly with human perception, while distance measures computed over discretized self-supervised speech tokens correlate substantially better. source: §4.1, Table 1 evidence: Averaged Pearson correlation with human PMOS ratings is r-bar = 0.30 for log F0 RMSE and r-bar = 0.66 for MCD, versus r-bar = 0.77 for the proposed token-edit-distance metric (DS-WED), aggregated via Fisher's Z transformation across 1000 samples and 2000 ratings. confidence: high relevance: high - claim_id: ar_generation_more_prosodically_diverse_than_flow_matching_but_not_mgm role: supports claim: Autoregressive generation provides greater output diversity in prosody than non-autoregressive flow-matching models with implicit text-speech alignment, but this advantage does not extend to non-autoregressive masked generative modeling. source: §4.4-4.5, Table 2 evidence: On the DS-WED benchmark across LibriSpeech test-clean and Seed-TTS test-en, three AR systems (XTTS-v2, CosyVoice, CosyVoice 2) outperform three flow-matching NAR systems (E2 TTS, F5-TTS, ZipVoice), while the masked generative modeling system (MaskGCT) surpasses all AR systems on LibriSpeech and remains competitive on Seed-TTS despite training on the same Emilia corpus as the flow-matching systems. confidence: high relevance: low - claim_id: duration_control_significant_but_insufficient_for_nar_prosody_diversity role: complicates claim: Explicit duration control during inference is a significant, but not sufficient, factor in restoring prosodic diversity to non-autoregressive TTS systems with implicit alignment. source: §4.5, Table 3 evidence: Applying duration perturbation (0.8-1.2x) to two flow-matching/MGM NAR systems increases DS-WED diversity by 13.8-28.5%, but the perturbed flow-matching system (F5-TTS) still lags behind AR and MGM systems evaluated without perturbation. confidence: medium relevance: low - claim_id: intelligibility_dpo_suppresses_prosody_diversity role: complicates claim: Preference optimization that targets one quality dimension (intelligibility) can measurably suppress output diversity along an unrelated axis (prosody) as a side effect. source: §4.5, Table 4 evidence: Applying DPO for intelligibility to CosyVoice 2 and MaskGCT reduces DS-WED prosody-diversity scores by 18.8% and 2.9-3.2% respectively across two test sets, with no diversity-specific reward term in the alignment objective. confidence: medium relevance: high - claim_id: llm_judge_unreliable_for_finegrained_prosodic_variation role: contradicts claim: A general-purpose large audio language model with strong multimodal reasoning capability can serve as a reliable automatic judge of fine-grained prosodic variation. source: §4.5, Table 5 evidence: Gemini 2.5 Pro, prompted to rate relative prosodic difference within groups of five samples, achieves only a weak correlation with human PMOS ratings (r-bar = 0.27) and unstable, wide-confidence-interval correlations with objective acoustic and token-based metrics. confidence: high relevance: high limitations: - DS-WED has only been validated on English speech; the benchmark covers only seven open-source systems and two evaluation corpora, and the DPO and duration-perturbation exploration each cover only two systems, limiting generalization of the paradigm-level conclusions. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2505.10599' published_date: "2025-09-25" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - autoregressive-LM - flow-matching - GAN - hybrid relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - autoregressive_generation - gan_adversarial_generation - flow_matching_generation - hybrid_generation claims: - claim_id: combining_discrete_emotion_labels_and_continuous_dimensional_emotion role: supports claim: Combining discrete emotion labels and continuous dimensional emotion representations within one conditioning scheme, rather than choosing one, expands the practically controllable range of emotional TTS beyond what either representation supports alone. source: §4.4, §4.6, Figure 4 evidence: Semi-supervised training across label-only, ADV-only, and jointly-annotated datasets raises ADV-space controllable coverage from 70% (training on jointly-annotated data alone) to 89.35% (full semi-supervised training), including synthesis of previously unseen ADV combinations. confidence: high relevance: high - claim_id: the_reliability_of_continuous_emotion_control_depends_heavily role: complicates claim: The reliability of continuous emotion control depends heavily on how the continuous space is quantized into controllable units, not just on the underlying representation. source: §4.4, Table 2, §4.6 evidence: Replacing the clustering-based nonlinear ADV quantizer with linear binning drops Spearman's Rank Correlation for perceived linear emotion control across all three ADV dimensions (e.g., valence SRC falls from 0.92 to 0.57), showing that imbalanced raw annotation distributions bias the model toward overrepresented regions of the space unless the quantizer compensates. confidence: high relevance: high - claim_id: predicting_an_emotion_conditioning_signal_directly_from_input role: supports claim: Predicting an emotion conditioning signal directly from input text allows emotional TTS to operate without an explicit emotion label or reference at inference time. source: §4.5, §4.6, Table 3 evidence: An ADV predictor infers pseudo-ADV tokens from text alone; removing it biases synthesis toward neutral emotion and drops pairwise preference against baselines from 67.33%/58.60% to 46.88%/28.50% in end-to-end preference tests. confidence: high relevance: high - claim_id: fine_grained_control_over_continuous_emotion_representations_remains role: complicates claim: Fine-grained control over continuous emotion representations remains bounded by the consistency of the human annotations used to train the control mechanism. source: §Limitations evidence: The authors note that subjective inter-annotator variation in ADV labels degrades linear control accuracy, and that texts with ambiguous emotional attributes cause the ADV predictor to infer inappropriate values, since the same text can express different emotions in different contexts. confidence: high relevance: high limitations: - Training and evaluation are restricted to English-language datasets; the paper does not report whether the semi-supervised ADV/label fusion strategy transfers to multilingual emotional speech. The ADV predictor's text-only inference mode is explicitly noted to fail on texts with ambiguous emotional attributes, since a single sentence can carry different emotions depending on context that the model cannot observe. The framework also depends on the quality and consistency of ADV annotations, which the authors identify as inherently noisy due to subjective variation across human annotators; they suggest more data and more consistent annotation as the primary mitigation rather than a modeling fix. Speaker conditioning uses a fixed embedding averaged from a speaker's neutral-emotion utterances rather than zero-shot cloning from an arbitrary reference clip, which is a narrower speaker-control setting than several of the LLM-based baselines it is compared against. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.20802' published_date: "2025-09-25" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - autoregressive-LM relevance: medium evidence_role: - core_evidence current_role: active_evidence method_family: - autoregressive_generation claims: - claim_id: many_transformer_layers_in_autoregressive_llm_tts_backbones role: supports claim: Many transformer layers in autoregressive LLM-TTS backbones contribute little to synthesis quality and can be removed with minimal loss in naturalness and speaker similarity. source: §4.1, Table 1 evidence: Halving CosyVoice 2's depth to 12 layers (39.7% fewer parameters) increases Seed-TTS WER by only 0.68 and decreases NMOS by 0.13, while speaker similarity and UTMOS remain essentially unchanged. confidence: high relevance: high - claim_id: layer_importance_criteria_developed_for_pruning_text_only role: refines claim: Layer-importance criteria developed for pruning text-only LLMs do not transfer directly to speech generation backbones; intelligibility-grounded criteria are needed to identify prunable layers correctly. source: §2.1, §4.2, Table 2 evidence: Cosine-based layer importance (input/output latent similarity) diverges from WER-based importance in TTS backbones, and substituting it for the proposed WER-based criterion increases WER from 1.59 to 1.74 and CER from 0.54 to 0.61 on LibriTTS test-clean. confidence: high relevance: low - claim_id: knowledge_distillation_can_recover_most_of_the_performance role: supports claim: Knowledge distillation can recover most of the performance lost from aggressive layer pruning in speech-generation LLMs using only a small fraction of the original pretraining data. source: §3, §4.1, Table 1b evidence: Fine-tuning pruned variants required under 5% of the original pretraining data (25% of LibriTTS for CosyVoice 2, an upper-bounded 12.5% of LibriHeavy for LLaSA) yet speaker similarity and UTMOS changed by at most 0.045 and 0.04 respectively relative to the uncompressed backbones. confidence: high relevance: high - claim_id: robustness_to_layer_pruning_varies_substantially_across_llm role: complicates claim: Robustness to layer pruning varies substantially across LLM-TTS backbones depending on how redundant their transformer layers are, so a single pruning ratio does not generalize uniformly. source: §4.1 evidence: LLaSA showed a larger relative quality drop after 50% layer pruning (speaker similarity −0.045, UTMOS −0.04) than CosyVoice 2 at the same pruning ratio, attributed to LLaSA's WLI values being more uniformly high across layers, indicating less exploitable redundancy. confidence: high relevance: high limitations: - The evaluation covers only two backbones (CosyVoice 2 and LLaSA-1B) and English-only test sets (LibriTTS test-clean, Seed-TTS test-en), so it is untested whether the WER-based pruning criterion and dynamic distillation scheme generalize to other LLM-TTS architectures, multilingual settings, or streaming inference. The more aggressive pruning configurations (9-layer CosyVoice 2) trade a larger, unquantified increase in WER for additional speed and memory gains, and the paper does not characterize where this trade-off becomes unacceptable for deployment. Computing WLI itself requires running WER evaluation over a data subset for each candidate layer removal, adding an upfront cost to the pruning procedure that is not reported in wall-clock terms. caveats: - For subjective evaluation, this paper provides medium-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2506.21875' published_date: "2025-09-26" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - SCA - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: s2t_performance_does_not_predict_s2s_quality role: supports claim: Strong performance on speech-to-text dialogue benchmarks does not reliably predict a model's speech-to-speech generation quality. source: §3.2, Appendix G evidence: Kimi-Audio's text-response quality on VoiceBench's CommonEval subset drops from 3.97 (text-only mode) to 3.00 (dual-modal speech+text mode), and its overall WildSpeech-Bench score (4.54) trails several models it is known to outperform on S2T-only benchmarks. confidence: high relevance: low - claim_id: e2e_speech_llms_remain_vulnerable_to_background_noise role: complicates claim: End-to-end speech LLMs remain broadly vulnerable to background acoustic interference despite strong performance in clean conditions. source: §3.4, Figure 5 evidence: Adding either environmental noise (ESC-50) or human background speech (LibriTTS) causes a significant performance decline across every evaluated model, including the top-performing GPT-4o-Audio, with human background speech producing the larger drop. confidence: high relevance: high - claim_id: cascaded_pipelines_match_e2e_except_on_paralinguistic_tasks role: refines claim: Cascaded ASR-LLM-TTS pipelines can match dedicated end-to-end speech LLMs on general conversational quality, but the gap re-emerges specifically on tasks requiring paralinguistic understanding. source: §3.2, Table 2 evidence: The naive Whisper + Qwen2.5-7B + MeloTTS pipeline scores comparably to GLM-4-Voice and MiniCPM overall (5.24 vs. 5.03/5.08) but scores markedly worse on the stress (2.95) and tone (4.12) paralinguistic subcategories, consistent with prosodic information being lost during ASR transcription. confidence: medium relevance: low - claim_id: query_aware_checklist_evaluation_improves_judge_alignment_with_humans role: supports claim: Query-aware, checklist-based evaluation improves an LLM judge's alignment with human judgment over generic scoring rubrics for speech-to-speech systems. source: §3.3, Figure 4 evidence: An ablation removing either the query-aware checklist or the multi-round ASR-plus-scoring procedure reduces the Pearson correlation between automatic and human expert scores relative to the full evaluation method. confidence: high relevance: high limitations: - Automatic scoring depends on a proprietary judge (GPT-4o-mini) and a proprietary ASR model (Whisper-large-v3) in the evaluation loop; the benchmark is deliberately restricted to English, non-toxic, single-turn queries. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.24457' published_date: "2025-09-29" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - codec - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: neural_network_metrics_correlate_more_strongly_than_classical_dsp_metrics role: supports claim: Neural-network-based objective quality metrics correlate more strongly with human subjective judgments than classical DSP-based metrics when evaluating neural audio codec output. source: §4 evidence: Across 1,700 data points from 17 codec conditions, scoreq_ref reaches a Pearson correlation of 0.87 with MUSHRA-1S scores and utmos, nomad, scoreq_nr, sheet_ssqa, and audiobox ce all exceed 0.80, while the strongest classical baselines (warpq, pesq) reach only 0.73. confidence: high relevance: high - claim_id: nonintrusive_mos_trained_metrics_saturate_near_top_of_quality_scale role: complicates claim: Non-intrusive objective metrics trained on MOS-based subjective ratings saturate near the top of the quality scale, weakening their ability to discriminate between near-transparent codec conditions. source: §4 evidence: Spearman and Kendall correlations for utmos and scoreq_nr drop sharply relative to their Pearson correlation, and condition-wise plots show scoreq_nr, ce, sheet_ssqa, and utmos producing near-constant scores in the high-MUSHRA range, attributed to the coarse 5-point ACR scale underlying the MOS targets these metrics are trained on. confidence: high relevance: high - claim_id: intrusive_metrics_more_stable_than_nonintrusive_at_high_quality role: refines claim: Reference-based (intrusive) objective metrics remain more statistically stable and discriminative than reference-free (non-intrusive) metrics for high-quality neural codec conditions, even though non-intrusive metrics are the only option when no reference signal is available. source: §4 evidence: Confidence-interval analysis across a high-, medium-, and low-quality codec condition shows intrusive scoreq_ref producing consistently smaller confidence intervals than non-intrusive scoreq_nr across sample sizes, an effect most pronounced for the high-quality condition. confidence: high relevance: high - claim_id: reference_based_classical_metrics_can_also_fail_to_discriminate role: complicates claim: Reference-based classical metrics can also fail to discriminate between codec conditions that differ substantially in perceived subjective quality, despite having access to the clean reference signal. source: §4 evidence: pesq shows the widest observed blind spot among all evaluated metrics, mapping codecs spanning roughly 35 MUSHRA points to nearly identical objective scores in the low-to-medium quality range, a limitation the authors note is a previously known phenomenon for pesq. confidence: high relevance: high limitations: - All results are obtained under clean speech conditions only, with noise and reverberation explicitly left for future work; the study is restricted to 12 codecs and 17 conditions from a single test set of 100 files. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.25416' published_date: "2025-09-29" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - TTS architecture: - diffusion relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: - diffusion_generation claims: - claim_id: easpo_stepwise_beats_endpoint_dpo role: supports claim: Applying preference supervision at each denoising step of a diffusion model, rather than only at the final generated output, improves alignment with a fine-grained target attribute compared to endpoint-propagated preference optimization. source: §3.4, Table 8 evidence: Under a matched Grad-TTS backbone, the stepwise EASPO method outperforms three prior diffusion-DPO variants (DDPO, D3PO, Diffusion-DPO) on emotion similarity, prosody similarity, WER, and UTMOS. confidence: high relevance: high - claim_id: easpo_endpoint_propagation_assumption_invalid role: complicates claim: The assumption underlying trajectory-level preference propagation in diffusion models, that all intermediate states on a path toward a preferred final output are themselves preferred, does not reliably hold. source: §1, §2.2 evidence: The paper motivates its stepwise design specifically by rejecting this assumption, and shows local win/lose selection at a shared latent outperforms endpoint-propagated variants. confidence: medium relevance: high - claim_id: easpo_contrastive_scorer_on_noisy_states role: supports claim: A contrastive audio-language scoring model can be adapted to evaluate noisy intermediate diffusion states, enabling automatic preference-pair construction without categorical emotion classifiers. source: §2.1, Table 3 evidence: Adding a time-aware normalization layer to a CLAP-based encoder (CLEP) and training it with diffusion-perturbed audio pairs lets it score emotional-prompt consistency directly on noisy mel-spectrograms; removing this timestep conditioning degrades Emo SIM, Prosody SIM, WER, and UTMOS. confidence: high relevance: high - claim_id: easpo_topcandidate_rollout_bias role: complicates claim: In RLHF-style fine-tuning of diffusion generation, continuing rollouts from the highest-reward candidate at each step, rather than a randomly selected candidate, biases training toward degenerate regions and reduces overall quality. source: §2.1, §3.4, Table 4 evidence: The paper reports that random selection of the next rollout state from the candidate pool consistently improves performance relative to continuing from the top-ranked (winning) candidate. confidence: high relevance: low limitations: - 'Evaluation is confined to a single diffusion TTS backbone (Grad-TTS) and a single, small English emotion corpus (ESD: 5 emotions x 10 speakers, 350 utterances/emotion).' - The paper does not report parameter counts for the fine-tuned decoder or wall-clock/inference overhead introduced by the k-candidate sampling and reward scoring at each training step, which matters for RLHF-style diffusion fine-tuning cost. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. - id: '2509.26542' published_date: "2025-09-30" entry_date: '2026-07-27' year: 2025 venue: arXiv task: - SCA - evaluation architecture: [] relevance: high evidence_role: - core_evidence - evaluation_caution current_role: active_evidence method_family: [] claims: - claim_id: substantial_text_voice_reasoning_gap_widens_with_reasoning_steps role: supports claim: A substantial reasoning-accuracy gap exists between text and voice modalities within the same model family, and the gap widens with the number of reasoning steps a task requires. source: §5.1, Table 3 evidence: GPT-5 text reaches 74.8% accuracy on competition math versus 6.1% for GPT-realtime voice (a 68.7-point gap), while the same model pair differs by only ~21 points on factual recall (48.3% vs. 27.4%); macro-averaged across five tracks, text models reach 54.0% vs. 11.3% for voice. confidence: high relevance: low - claim_id: more_thinking_time_insufficient_to_close_voice_reasoning_gap role: complicates claim: Increasing inference-time deliberation ('thinking time') before a voice system speaks is not, by itself, sufficient to close a voice-native reasoning gap. source: §5.2 evidence: Enabling Audio Flamingo 3's thinking mode raises time-to-first-response by 530% (2.40s to 15.14s) but overall accuracy decreases from 1.7% to 1.5%, with Context-track accuracy dropping from 3.8% to 1.8%. confidence: high relevance: low - claim_id: reasoning_narration_decoupling_narrows_but_does_not_close_gap role: complicates claim: Decoupling reasoning from speech narration in a cascade architecture narrows a voice-native reasoning gap but introduces new grounding and logical-consistency failures that prevent full parity with text. source: §5.2, Figure 5 evidence: A cascade pairing a strong text reasoner with a separate narration model still trails the text-only upper bound by 15.7 points on Math (59.1% vs. 74.8%) and collapses to 0.2% accuracy on the Context track, with elevated rates of UNSUPPORTED_FACT and LOGICAL_CONTRADICTION errors attributed to the narration stage. confidence: high relevance: low - claim_id: speech_intelligibility_not_primary_bottleneck_for_voice_reasoning role: supports claim: Speech output intelligibility is not the primary bottleneck limiting reasoning performance in voice-interactive systems. source: §5.2, Table 3 evidence: Word Error Rate across evaluated voice systems ranges from 7.9% (Gemini-2.5-Flash-audio) to 19.8% (Freeze-Omni) with no corresponding difference in reasoning accuracy; all systems in this WER range perform uniformly poorly on VERA's reasoning tracks. confidence: high relevance: low - claim_id: voice_architectures_fail_reasoning_in_systematically_different_ways role: refines claim: Different voice system architectures fail on reasoning tasks in systematically different, architecture-linked ways rather than through a single uniform error pattern. source: §5.3, Figure 6 evidence: Native streaming models (GPT-realtime, Gemini-2.5-Flash-Audio) under-produce NO_FINAL_ANSWER and OFF_TARGET errors, favoring fluent-but-incorrect completions, while the LiveAnswer cascade over-produces UNSUPPORTED_FACT, OFF_TARGET, and LOGICAL_CONTRADICTION deviations, and end-to-end models (Moshi, Qwen2.5-Omni) show yet another bimodal pattern of extreme over- or under-completion. confidence: medium relevance: high limitations: - The study is an observational, cross-architecture diagnostic rather than a controlled experiment, since the 12 evaluated voice systems differ in training data, objectives, and proprietary implementation details, preventing isolation of modality as the sole causal factor. caveats: - For subjective evaluation, this paper provides high-relevance evidence; claims primarily concerning other concepts remain included but are marked low relevance. claim_clusters: - id: automatic_metrics_diverge_from_human_perception claim: Automatic speech-quality and task metrics frequently rank systems differently from human perceptual judgments. status: strongly_supported confidence: high supporting_papers: - '1609.03499' - '2402.01912' - '2403.16973' - '2406.18009' - '2506.09874' - 2025.icnlsp-1.34 - '2508.02013' - '2504.12867' - '2503.04721' - interspeech-2025-0383 - interspeech-2025-1084 - interspeech-2025-1434 - interspeech-2025-2573 - '2509.04093' - '2509.13989' - '2509.18470' - '2509.18531' - '2509.18823' - '2509.18928' - '2509.19928' - '2506.21875' - '2509.24457' contradicting_papers: [] refining_papers: - 2025.chipsal-1.18 - '2507.06235' - '2507.01611' - '2508.00317' - '2508.07711' - interspeech-2025-0438 - interspeech-2025-0469 - interspeech-2025-0779 - interspeech-2025-1084 - interspeech-2025-1434 - interspeech-2025-1478 - interspeech-2025-1639 - interspeech-2025-1726 - interspeech-2025-1747 - interspeech-2025-2787 - interspeech-2025-bokkahallisatish25_interspeech - '2508.17623' - '2509.00675' - '2506.23367' - '2509.05863' - '2509.06074' - '2509.19812' - '2509.24457' - '2509.19928' caveats: - The direction and size of disagreement depend on task, metric, quality range, and evaluation population. last_reviewed: '2026-07-27' - id: learned_quality_metrics_outperform_classical_metrics claim: Learned perceptual-quality models generally correlate with human ratings better than classical signal-based metrics, but remain domain- and scale-sensitive. status: strongly_supported confidence: high supporting_papers: - '2204.02152' - '2507.06116' - '2508.00317' - '2508.08957' - interspeech-2025-0973 - interspeech-2025-0984 - '2509.03292' - '2509.15629' - '2509.18823' - '2509.24457' contradicting_papers: [] refining_papers: - '2204.02152' - '2507.01611' - '2507.06116' - interspeech-2025-1747 - '2509.18823' - '2509.24457' caveats: - High aggregate correlation can conceal utterance-level errors, domain shift, and saturation near transparent quality. last_reviewed: '2026-07-27' - id: mos_protocols_are_context_and_population_sensitive claim: MOS-style evaluations are sensitive to listener population, cultural context, reference conditions, and protocol design. status: strongly_supported confidence: medium supporting_papers: - '2204.02152' - 2025.americasnlp-1.1 - '2505.15772' - interspeech-2025-0762 contradicting_papers: [] refining_papers: - 2025.naacl-long.464 - 2025.naacl-short.69 - 2025.americasnlp-1.1 - '2505.15772' - '2508.07711' caveats: - MOS remains useful for broad naturalness comparisons when sampling, anchors, instructions, and listener recruitment are reported rigorously. last_reviewed: '2026-07-27' - id: pairwise_and_ranking_designs_increase_discrimination claim: Pairwise preference, ranking, and contrastive evaluation designs often discriminate perceptual differences more reliably than isolated absolute scores. status: strongly_supported confidence: high supporting_papers: - '2105.06337' - '2204.02152' - '2406.00654' - 2025.coling-main.352 - 2025.findings-acl.1226 - 2025.sigdial-1.21 - '2508.02013' - '2508.08957' - interspeech-2025-0383 - interspeech-2025-1550 - interspeech-2025-bokkahallisatish25_interspeech - '2508.15931' - '2509.00685' - '2509.04093' - '2505.10599' contradicting_papers: [] refining_papers: - '2506.09874' - 2025.acl-long.388 - interspeech-2025-2765 caveats: - Pairwise outcomes are relative to the selected baseline and do not by themselves establish absolute quality. last_reviewed: '2026-07-27' - id: preference_optimization_improves_targeted_perception claim: Preference-based optimization can improve targeted perceptual attributes beyond supervised or metric-only training. status: strongly_supported confidence: high supporting_papers: - '2406.00654' - 2025.acl-long.598 - 2025.acl-long.997 - interspeech-2025-2189 - '2509.00685' - '2509.05863' - '2509.14946' - '2509.18531' - '2509.18928' - '2509.25416' contradicting_papers: [] refining_papers: - 2025.acl-long.598 - 2025.acl-long.997 - '2509.05863' caveats: - Reported gains often target one perceptual axis and may rely on proxy-generated rather than human-labeled preferences. last_reviewed: '2026-07-27' - id: preference_alignment_has_cross_dimension_tradeoffs claim: Optimizing one perceptual or proxy objective can degrade untargeted dimensions such as naturalness, identity, or prosodic diversity. status: strongly_supported confidence: medium supporting_papers: - '1609.03499' - '2507.00808' - interspeech-2025-1397 - interspeech-2025-2573 - '2509.15629' contradicting_papers: [] refining_papers: - '2507.04817' - 2025.icnlsp-1.34 - '2509.05863' - '2509.14684' - '2509.16010' - '2509.18531' - '2509.18928' - '2509.19928' caveats: - Trade-offs depend on reward construction, reference policy, stopping criteria, and the representational ceiling of the generator. last_reviewed: '2026-07-27' - id: speaker_similarity_metrics_are_confounded claim: Embedding-based speaker-similarity metrics are confounded by recording, architecture, language, and reference design and can disagree with perceived identity. status: emerging confidence: medium supporting_papers: - '2507.20140' contradicting_papers: [] refining_papers: - '2507.02176' - interspeech-2025-0438 - interspeech-2025-1081 - interspeech-2025-1084 - interspeech-2025-1726 - '2509.05863' - '2412.16846' - '2509.15629' caveats: - Speaker-verification metrics remain useful within tightly controlled protocols but should not be treated as direct substitutes for listener identity judgments. last_reviewed: '2026-07-27' - id: fine_grained_style_and_prosody_need_dimension_specific_tests claim: Fine-grained prosody, emotion, accent, and style control require dimension-specific perceptual tests rather than holistic naturalness scores alone. status: strongly_supported confidence: high supporting_papers: - '1609.03499' - '2406.00654' - 2025.coling-main.352 - 2025.iwsds-1.11 - 2025.iwsds-1.27 - '2505.15772' - '2507.04817' - 2025.acl-long.388 - 2025.sigdial-1.27 - '2507.20091' - '2504.12867' - interspeech-2025-0464 - interspeech-2025-0648 - interspeech-2025-0706 - interspeech-2025-0762 - interspeech-2025-0902 - interspeech-2025-1098 - interspeech-2025-1115 - interspeech-2025-1210 - interspeech-2025-1394 - interspeech-2025-1638 - interspeech-2025-1639 - interspeech-2025-1684 - interspeech-2025-1940 - interspeech-2025-2189 - interspeech-2025-2283 - interspeech-2025-2586 - interspeech-2025-2765 - '2508.13028' - '2508.17494' - '2508.17623' - '2509.00685' - '2509.03940' - '2509.04093' - '2509.06074' - '2509.09716' - '2508.18240' - '2509.14946' - '2509.15626' - '2509.20378' - '2509.19231' - '2509.19928' - '2505.10599' - '2509.25416' contradicting_papers: [] refining_papers: - 2025.coling-main.352 - 2025.iwsds-1.11 - '2507.06235' - '2505.15772' - '2507.04817' - '2507.14988' - '2507.16875' - 2025.conll-1.9 - 2025.sigdial-1.27 - interspeech-2025-0433 - interspeech-2025-0854 - interspeech-2025-1098 - interspeech-2025-1210 - interspeech-2025-1639 - interspeech-2025-1940 - interspeech-2025-2595 - interspeech-2025-2787 - '2508.13028' - '2508.17494' - '2508.17623' - '2506.23367' - '2509.06074' - '2508.18240' - '2509.14946' - '2509.16589' - '2509.18531' - '2509.19928' - '2505.10599' caveats: - Attribute-specific tests improve diagnosticity but require validated labels, anchors, and culturally appropriate definitions. last_reviewed: '2026-07-27' - id: quality_control_tradeoffs_are_perceptually_nonmonotonic claim: Greater control strength, compression, guidance, or transformation does not produce monotonic perceptual gains and often exposes quality trade-offs. status: strongly_supported confidence: high supporting_papers: - '1609.03499' - '2105.06337' - interspeech-2025-0464 - interspeech-2025-1084 - interspeech-2025-1122 - interspeech-2025-2586 - '2509.08696' - '2509.17143' contradicting_papers: [] refining_papers: - 2025.coling-main.352 - '2507.04817' - '2507.16875' - interspeech-2025-0383 - interspeech-2025-1531 - interspeech-2025-1747 - interspeech-2025-1940 - interspeech-2025-2032 - interspeech-2025-2595 - '2509.04667' - '2509.04685' - '2509.05863' - '2509.07038' - '2505.17093' caveats: - The optimum is system- and attribute-specific, so evaluations should report curves rather than a single operating point. last_reviewed: '2026-07-27' - id: evaluation_must_cover_domain_and_language_shift claim: Subjective conclusions from clean, high-resource, or single-language evaluations often fail to transfer to low-resource and in-the-wild conditions. status: strongly_supported confidence: high supporting_papers: - '2204.02152' - '2206.04658' - '2403.16973' - 2025.chipsal-1.18 - 2025.computel-main.6 - 2025.naacl-short.69 - '2507.16875' - 2025.acl-long.598 - '2508.00317' - '2508.08715' - interspeech-2025-0469 - interspeech-2025-0706 - interspeech-2025-0854 - interspeech-2025-2151 - '2508.13028' - '2509.05863' - '2509.17988' contradicting_papers: [] refining_papers: - '2204.02152' - '2407.05361' - 2025.chipsal-1.18 - 2025.naacl-long.464 - 2025.naacl-long.619 - interspeech-2025-0305 - interspeech-2025-0854 - interspeech-2025-1531 - '2509.05863' caveats: - Cross-domain comparisons are difficult when datasets, listener groups, and recording conditions change simultaneously. last_reviewed: '2026-07-27' - id: demographic_and_social_bias_affect_perceived_quality claim: Demographic representation and social perceptions can create systematic disparities in subjective speech-system quality and identity judgments. status: strongly_supported confidence: high supporting_papers: - '1904.02882' - '2507.06235' - 2025.acl-long.1252 - interspeech-2025-2031 - interspeech-2025-2573 - interspeech-2025-bokkahallisatish25_interspeech contradicting_papers: [] refining_papers: - interspeech-2025-0383 - interspeech-2025-1081 - interspeech-2025-bokkahallisatish25_interspeech - '2509.07376' - '2509.13989' caveats: - Observed disparities may combine training-data imbalance, evaluator composition, social stereotypes, and acoustic confounds. last_reviewed: '2026-07-27' - id: participatory_evaluation_adds_context_beyond_scores claim: Participatory and qualitative evaluation reveals usability, cultural, and pedagogical dimensions that scalar listener scores omit. status: strongly_supported confidence: medium supporting_papers: - 2025.computel-main.6 - '2503.04721' - interspeech-2025-bokkahallisatish25_interspeech - '2509.08696' contradicting_papers: [] refining_papers: - 2025.computel-main.6 caveats: - Small community studies prioritize contextual validity over population-level statistical generalization. last_reviewed: '2026-07-27' - id: llm_and_audio_language_judges_are_not_uniformly_reliable claim: Automated LLM and audio-language-model judges can approximate some human speech judgments but are not uniformly reliable for fine-grained or non-verbal attributes. status: contested confidence: medium supporting_papers: - '2508.02013' - '2504.12867' - interspeech-2025-0706 - '2509.03940' - '2509.09716' - '2506.21875' contradicting_papers: - '2509.19928' refining_papers: - '2508.18240' - '2509.16589' caveats: - Reliability varies by input modality, rubric specificity, repeated sampling, and whether the judged attribute is verbal or paralinguistic. last_reviewed: '2026-07-27' - id: benchmark_design_needs_multidimensional_coverage claim: Speech-system benchmarks need multidimensional rubrics that separate semantic success, naturalness, identity, prosody, and interaction behavior. status: strongly_supported confidence: high supporting_papers: - '2506.23049' - 2025.acl-long.911 - '2508.02013' - '2503.04721' - interspeech-2025-0739 - interspeech-2025-2151 - interspeech-2025-bokkahallisatish25_interspeech - '2508.17623' - '2509.04093' - '2509.09716' - '2508.18240' - '2506.21875' contradicting_papers: [] refining_papers: - '2407.05361' - 2025.chipsal-1.18 - '2505.15772' - 2025.acl-long.388 - interspeech-2025-0406 - interspeech-2025-2765 - '2508.17494' - '2509.04072' - '2509.26542' - interspeech-2025-0984 caveats: - Composite benchmarks can obscure dimension-level failure unless component scores and judge reliability are reported. last_reviewed: '2026-07-27' - id: perceptual_tests_validate_quality_speed_tradeoffs claim: Human listening tests are necessary to establish whether efficiency gains preserve perceived quality across synthesis architectures. status: strongly_supported confidence: high supporting_papers: - '1609.03499' - '2010.05646' - '2105.06337' - '2210.13438' - '2306.00814' - '2507.20731' - '2503.04721' - interspeech-2025-0554 - interspeech-2025-1081 - interspeech-2025-1084 - interspeech-2025-1747 - '2509.08696' - '2412.16846' - '2509.18470' - '2509.20802' contradicting_papers: [] refining_papers: - interspeech-2025-0739 - interspeech-2025-1364 - interspeech-2025-1747 - '2509.08696' - '2509.20802' caveats: - Many comparisons vary model size, hardware, data, or decoder jointly with speed, limiting causal attribution. last_reviewed: '2026-07-27' - id: codec_and_vocoder_choices_set_perceptual_ceilings claim: Codec and vocoder design choices impose perceptual ceilings and artifacts that propagate into downstream speech-generation evaluations. status: strongly_supported confidence: high supporting_papers: - '1712.05884' - '2010.05646' - '2206.04658' - '2210.13438' - '2301.11325' - '2306.00814' - '2402.01912' - '2409.05377' - 2025.chipsal-1.18 - '2507.01611' - '2507.20731' - '2508.07711' - '2508.08715' - interspeech-2025-1122 - interspeech-2025-1625 - interspeech-2025-1639 - interspeech-2025-1763 - interspeech-2025-1819 - interspeech-2025-2032 - '2509.18823' - '2509.24457' contradicting_papers: [] refining_papers: - '2508.01796' - interspeech-2025-1763 - interspeech-2025-2032 - interspeech-2025-2739 - '2509.05863' - '2509.18823' caveats: - Codec effects interact with sample rate, bitrate, training targets, and whether evaluation uses reconstructed references. last_reviewed: '2026-07-27' - id: naturalness_intelligibility_and_identity_are_distinct_axes claim: Naturalness, intelligibility, speaker identity, and expressiveness are distinct perceptual axes and improvements on one do not guarantee gains on the others. status: strongly_supported confidence: high supporting_papers: - '2406.18009' - '2409.05377' - 2025.computel-main.6 - 2025.americasnlp-1.1 - '2507.04817' - '2507.09310' - '2507.14988' - 2025.acl-long.598 - '2507.20140' - '2508.06870' - '2508.09702' - interspeech-2025-0438 - interspeech-2025-1434 - interspeech-2025-2151 - '2509.00685' - '2509.15629' - '2509.17143' - '2509.18928' - '2509.20802' contradicting_papers: [] refining_papers: - 2025.americasnlp-1.1 - '2507.04817' - '2507.16875' - 2025.icnlsp-1.34 - interspeech-2025-0438 - interspeech-2025-1229 - interspeech-2025-1478 - interspeech-2025-2595 - '2509.04667' - '2509.05863' - '2509.16010' - '2509.17143' - '2509.18928' - '2509.20802' caveats: - Correlations vary with task and quality range; evaluations should report each axis separately rather than collapse them into one score. last_reviewed: '2026-07-27' method_families: - id: autoregressive_generation name: Autoregressive generation summary: Autoregressive language-model and waveform systems whose perceptual evaluation must account for sequential generation, exposure bias, codec ceilings, and sampling behaviour. papers: - '1609.03499' - '2301.11325' - '2305.07243' - '2402.01912' - '2403.16973' - '2406.00654' - '2407.08551' - '2506.23049' - 2025.acl-long.598 - 2025.acl-long.997 - 2025.icnlsp-1.34 - '2507.20091' - '2508.08715' - '2504.12867' - interspeech-2025-0464 - interspeech-2025-1084 - interspeech-2025-1115 - interspeech-2025-2787 - '2509.00685' - 2025.clicit-1.27 - '2509.04072' - '2509.04093' - '2509.05863' - '2412.16846' - '2509.14946' - '2509.20378' - '2509.17021' - '2509.17143' - '2509.18531' - '2509.18928' - '2505.10599' - '2509.20802' open_questions: - How should listening tests separate representation limits from autoregressive decoding and sampling effects? - id: transformer_encoder_decoder name: Transformer encoder–decoder synthesis summary: Sequence-to-sequence and non-autoregressive transformer systems evaluated for alignment, naturalness, intelligibility, and controllable prosody. papers: - '1703.10135' - '1712.05884' - '2105.06337' - 2025.chipsal-1.18 - 2025.coling-main.352 - '2409.20007' - 2025.naacl-long.619 - 2025.americasnlp-1.1 - '2507.00808' - 2025.acl-long.388 - 2025.acl-long.911 - 2025.sigdial-1.27 - '2508.05385' - '2508.06870' - interspeech-2025-0305 - interspeech-2025-0383 - interspeech-2025-0408 - interspeech-2025-0469 - interspeech-2025-0854 - interspeech-2025-1098 - interspeech-2025-1122 - interspeech-2025-1394 - interspeech-2025-1550 - interspeech-2025-1940 - interspeech-2025-2031 - interspeech-2025-2189 - interspeech-2025-2586 - '2508.13028' - '2508.15931' - '2508.17031' - '2509.00675' - '2509.06074' open_questions: - Which evaluation protocols best expose alignment failures that aggregate MOS obscures? - id: gan_adversarial_generation name: GAN and adversarial generation summary: Adversarial vocoders, codecs, conversion systems, and hybrid evaluators where discriminator choices and waveform artefacts directly shape perceived quality. papers: - '2010.05646' - '2206.04658' - '2210.13438' - '2306.00814' - '2409.05377' - 2025.chipsal-1.18 - 2025.americasnlp-1.1 - '2507.01611' - '2507.04817' - '2507.20731' - '2508.01796' - '2508.06870' - '2508.07711' - '2508.08715' - interspeech-2025-0406 - interspeech-2025-0408 - interspeech-2025-0554 - interspeech-2025-0854 - interspeech-2025-0902 - interspeech-2025-1081 - interspeech-2025-1122 - interspeech-2025-1531 - interspeech-2025-1550 - interspeech-2025-1625 - interspeech-2025-1639 - interspeech-2025-1747 - interspeech-2025-1763 - interspeech-2025-1819 - interspeech-2025-2151 - interspeech-2025-2573 - interspeech-2025-2586 - interspeech-2025-2739 - '2508.17031' - '2509.04667' - '2509.04685' - '2509.18823' - '2509.19231' - '2509.19812' - '2505.10599' open_questions: - How consistently do modern listening protocols detect adversarial and codec artefacts across domains? - id: diffusion_generation name: Diffusion generation summary: Score-based and diffusion speech systems evaluated across denoising steps, guidance settings, preference objectives, and quality–latency trade-offs. papers: - '2105.06337' - '2305.07243' - 2025.coling-main.352 - 2025.naacl-long.619 - '2508.01796' - interspeech-2025-1066 - interspeech-2025-1115 - interspeech-2025-1210 - interspeech-2025-1364 - interspeech-2025-1397 - interspeech-2025-1434 - interspeech-2025-1747 - interspeech-2025-1763 - '2509.07038' - '2509.18470' - '2509.18928' - '2509.19231' - '2509.25416' open_questions: - Can diffusion evaluations control model capacity and inference budget while preserving blinded perceptual comparisons? - id: flow_matching_generation name: Flow-matching generation summary: Flow-matching acoustic and waveform systems whose evaluations span naturalness, alignment, prompt robustness, controllability, and low-step inference. papers: - '2406.18009' - 2025.naacl-short.69 - '2506.09874' - '2507.14988' - '2507.16875' - 2025.acl-long.313 - 2025.acl-long.598 - '2507.20140' - '2508.08715' - '2504.12867' - interspeech-2025-0305 - interspeech-2025-0554 - interspeech-2025-0762 - interspeech-2025-0854 - interspeech-2025-1229 - interspeech-2025-1940 - interspeech-2025-2573 - '2506.23367' - '2509.04072' - '2509.04093' - '2509.08696' - '2509.14684' - '2509.14946' - '2509.17988' - '2509.18470' - '2505.10599' open_questions: - Which perceptual differences between flow matching and diffusion remain under matched backbones and compute? - id: variational_latent_generation name: Variational and latent generation summary: VAE and continuous-latent systems assessed for diversity, reconstruction quality, disentanglement, and listener-perceived identity or style. papers: - '2305.07243' - '2409.05377' - '2507.09310' - 2025.acl-long.937 - interspeech-2025-0433 - interspeech-2025-0902 - interspeech-2025-1531 - interspeech-2025-1550 - '2509.07376' - '2412.16846' - '2509.15626' open_questions: - How should subjective studies distinguish latent diversity from instability or identity drift? - id: hybrid_generation name: Hybrid generation summary: Systems combining multiple synthesis paradigms or signal-processing stages, often exposing trade-offs that single automatic metrics fail to capture. papers: - 2025.acl-long.1252 - 2025.acl-long.598 - 2025.findings-acl.1226 - 2025.sigdial-1.21 - interspeech-2025-0438 - interspeech-2025-0656 - interspeech-2025-1478 - interspeech-2025-1638 - interspeech-2025-1684 - interspeech-2025-1726 - interspeech-2025-1776 - interspeech-2025-2032 - '2508.17494' - '2509.03292' - '2509.04667' - '2509.16010' - '2509.18928' - '2509.19883' - '2505.10599' open_questions: - How should evaluations attribute perceptual gains and losses to individual stages in hybrid pipelines? reassessment_queue: [] open_questions: - Which subjective protocol best balances ecological validity, diagnostic resolution, cost, and cross-laboratory reproducibility? - How should listener demographics, language competence, hearing profile, and cultural context be reported and incorporated into aggregate judgments? - When can learned quality predictors or LLM judges safely reduce human evaluation, and which attributes must remain human-rated? - How can preference optimization improve a target attribute without suppressing naturalness, identity, diversity, or other unmeasured dimensions? - Which reference and anchoring designs make speaker identity, prosody, and expressive-style judgments comparable across systems and datasets? trend_notes: - Subjective evaluation has expanded from aggregate naturalness MOS toward multidimensional tests of identity, prosody, emotion, interaction, and social perception. - Pairwise preferences and preference-guided training became prominent in 2024–2025, often using automatic proxies but increasingly exposing cross-dimension trade-offs. - Learned MOS predictors, codec embeddings, and audio-language-model judges are increasingly used to scale evaluation, while multiple studies document domain shift, saturation, and fine-grained reliability failures. - Low-resource and participatory studies increasingly adapt evaluation protocols to native-listener availability and cultural norms rather than importing standard MOS unchanged. - Benchmark design is moving toward hierarchical rubrics, arena-style comparisons, and query-specific checklists for spoken dialogue and instruction-conditioned systems.