arXiv · 2026 · Preprint
Koriagin et al. (T-Tech / AI Foundation and Algorithm Lab) · → Paper · Demo: ? · Code: ?
Trains BatchTopK sparse autoencoders on the residual stream of the frozen CosyVoice3 LM backbone, labels the resulting features by whether they fire on text-prefix or speech-token positions, and shows that shifting single SAE features during generation changes laughter, perceived speaker gender, and speech rate.
Problem
LM-based TTS systems process one residual stream that holds both a text/instruction prefix and the generated discrete speech tokens, yet there is no principled way to tell which internal directions encode linguistic content and which encode acoustic or prosodic properties of the speech being generated. Sparse autoencoders (SAEs) and LLM-based automatic interpretation (“auto-interp”) are established tools for text LMs, but their evidence formats (token windows) do not carry over directly to features that fire on speech tokens, and it is unclear whether such features are causally usable as control directions for synthesis.
Method
The object of study is CosyVoice 3, whose LM is a Qwen2.5-0.5B backbone (hidden size 896) that reads a BPE-tokenised text prompt and autoregressively generates 25 Hz discrete speech tokens, which are then rendered by a DiT flow-matching model and a HiFi-GAN vocoder (Figure 5). The whole TTS system stays frozen; the flow-matching decoder and vocoder are neither analysed nor steered. Only the SAEs are trained.
SAE training. BatchTopK SAEs with a 16,384-feature dictionary and k=50 active features per token are trained at multiple LM layers on roughly 250M tokens of Emilia, using the standard reconstruction plus sparsity objective with an auxiliary dead-feature loss (§3.1). Layer 20 is used for the detailed qualitative and steering experiments.
Modality-aware evidence and tagging. The LM is run teacher-forced on the layout [sos | instruct | text | task | speech], and exact segment lengths are recorded so each top activation can be assigned to the prefix or to the speech segment. For each feature, the fraction of its top-20 activations falling in the speech segment determines its modality: audio-modal (speech fraction at least 0.8), text-modal (at most 0.2), or mixed (§3.3, Appendix B). Text activations are shown to the labeller as marked token windows; speech activations are converted to 1-second audio clips centred on the active token via the 25 Hz token rate (§3.2).
Labelling and scoring. Gemini 3.0 Pro writes a one-sentence label from modality-matched evidence (text only, audio clips only, or both). A detection-style scorer, using the same Gemini model, then rates held-out positive evidence (activations ranked below the five used for labelling) and negatives drawn from other features on a 0-10 scale; AUROC and balanced accuracy (threshold at 5) are reported (§3.4-3.5, Appendix B).
SAE latent steering. At the chosen layer, speech-token residual states (the speech-prompt segment at prefill and the current generated token during decoding) are encoded by the frozen SAE, selected feature activations are shifted by α times a sign vector times feature-wise activation scales, and the modified latent is decoded back to replace the residual vector. Instruction, text and task tokens are untouched (§3.6, Appendix C).
Probing. Logistic-regression probes on raw residuals or SAE latents test linear decodability of laughter (VocalSound), four ESD emotions and eleven VCTK accents against an Emilia-Yodas neutral pool, with a top-k SAE coordinate re-fit (k=1 as a monosemanticity test) (Appendix G).
Key Results
Layer sweep (Figure 1, Appendix D). Reconstruction stays high in early layers (explained variance 0.97-0.99 at layers 0-8) and declines to a minimum of 0.82 at layer 20, rebounding to 0.945 at the final hidden state (layer 23). Feature modality shifts with depth: mixed features are most frequent in early and middle layers (peaking at 47.3% at layer 12), layers 16-20 are audio-heavy (76.1% audio-modal at layer 16, 74.3% at layer 20, with mixed features falling from 40.9% at layer 14 to 4.1% at layer 20), and layer 23 reverts to 83.1% text-modal. Text positions are reconstructed at least as well as audio positions in most layers, with the text-minus-audio EV gap at 0.065 (layer 16) and 0.080 (layer 20); at layer 4 audio EV is slightly higher (0.992 vs 0.986).
Auto-interp quality (Figure 2, Figure 6). At layer 20, on balanced 668-feature samples per modality, text-modal labels reach AUROC 0.921, audio-modal 0.653 and mixed 0.558. Across completed layers the same ordering holds (text 0.90-0.94, audio 0.65-0.72, mixed 0.53-0.69). Representative features include a text feature for “British” in accent descriptions, an audio feature for laughter (balanced accuracy and AUROC 0.75), audio phoneme features (/k/, /ing/, both 1.000/1.000) and mixed features linking stutters or the word “middle” across text and speech (Table 3). These per-feature scores are hand-picked examples, not averages.
Steering (Figure 4, Table 2). Averaged over a 40-voice by 10-text prompt grid:
| Feature (label) | Measure | Baseline (α=0) | Steered |
|---|---|---|---|
| 14834 (laughter) | laughter probability | 0.015 | 0.791 at α=+60 |
| 11402 (speaker gender) | wav2vec2 P(male) | 0.629 | 0.944 at α=-50; 0.063 at α=+50 |
| 3024 (speech rate) | voiced duration (s) | 3.96 | 10.57 at α=-50; 2.75 at α=+50 |
Figure 4 also shows off-target effects: the laughter feature raises voiced duration to roughly 10.7 s at α=+60, and the gender feature raises it to roughly 5.6 s at α=+50, while the speech-rate feature leaves P(male) near baseline. Split by prompt gender (Figure 3), male-prompted P(male) stays near 1.0 up to α=+25 before dropping sharply at α=+50, and female-prompted P(male) rises from about 0.28 to about 0.9 at α=-50. The text states that the speech-rate feature preserves spoken content, but no WER or other content metric is reported.
Probing (Tables 4-5). Raw-residual probes exceed 0.99 ROC-AUC for all three concepts by layer 8; SAE-latent laughter probes track them from layer 8. Single-coordinate (top-1) probes are weaker and layer-dependent: laughter peaks at 0.929 (L16) and falls to 0.695 at L20, emotion mean peaks at 0.928 (L16), accent mean at 0.925 (L8).
There are no baselines: steering is not compared against residual-vector steering or against prior TTS steering methods, and all steering metrics come from automatic classifiers.
Novelty Assessment
The SAE recipe (BatchTopK), the auto-interp protocol (LLM labeller plus detection scorer) and latent-space steering are all adapted from text-LM interpretability. The genuinely new elements are the application to a generative TTS LM backbone and the modality routing that converts speech-token activations into short audio clips so one labelling pipeline can handle text, audio and mixed features. The layer-wise modality profile (mixed, then audio-committed, then text-reverting at the final layer) is the most informative empirical finding. The steering results are a proof of concept on three hand-chosen features with no ablation of the steering design or comparison to alternatives, so the contribution is primarily an empirical interpretability analysis rather than a new control architecture.
Field Significance
moderate: The paper provides a layer-by-layer map of how a frozen LM-based TTS backbone splits its sparse features between text-prefix and speech-token positions, and demonstrates that individual SAE features at layer 20 can be shifted to change laughter, perceived gender and speech rate in the output. Its type is empirical-benchmark because the contribution is the analysis and its measurements on one existing model; the SAE and steering machinery are borrowed and not ablated, which rules out architectural-novelty. The evidence is limited to one 0.5B model, an LLM labeller that also scores its own labels, and classifier-only steering metrics.
Claims
- supports: Intermediate-to-late layers of an LM-based TTS backbone contain sparse features tied predominantly to generated speech-token positions, while the final layer reverts toward text-aligned features.
Evidence: Audio-modal features make up 76.1% at layer 16 and 74.3% at layer 20 of CosyVoice3’s Qwen2.5-0.5B backbone, mixed features fall from 40.9% (layer 14) to 4.1% (layer 20), and layer 23 is 83.1% text-modal. (§4.1, Figure 1b, Appendix D)
- supports: Sparse features in an LM-based TTS backbone can act as causal control directions for paralinguistic and prosodic properties without retraining the TTS model.
Evidence: Shifting layer-20 SAE features raises laughter probability from 0.015 to 0.791 (α=+60), moves wav2vec2 P(male) from 0.629 to 0.944 or 0.063 (α=-50 / +50), and changes voiced duration from 3.96 s to 10.57 s or 2.75 s (α=-50 / +50). (§4.4, Figure 4)
- complicates: Single-feature steering in a TTS LM is not attribute-isolated: steering one feature can shift other properties of the output.
Evidence: The laughter feature also lengthens voiced duration to roughly 10.7 s at α=+60 and the gender feature to roughly 5.6 s at α=+50; male-prompted gender shift appears only at the largest positive α. No content-preservation metric is reported for any steered condition. (§4.4, Figures 3-4)
- complicates: Automatically generated labels for speech-side features in TTS models are less reliable than labels for text-side features.
Evidence: At layer 20 the held-out detection AUROC is 0.921 for text-modal, 0.653 for audio-modal and 0.558 for mixed features, with the same ordering across completed layers; labeller and scorer share one Gemini model, which the authors flag as circular. (§4.2, Figure 2, Appendix E, §Limitations)
- refines: Speech-style concepts that are linearly decodable from a TTS backbone’s residual stream are only partly concentrated in single sparse features, and the degree of concentration varies by layer.
Evidence: Raw-residual probes exceed 0.99 ROC-AUC by layer 8 for laughter, emotion and accent, but top-1 SAE coordinate probes peak at 0.929 / 0.928 / 0.925 and drop to 0.695 for laughter at layer 20. (Appendix G, Tables 4-5)
Limitations and Open Questions
All steering evidence comes from automatic classifiers (laughter probability, a wav2vec2 gender classifier, voiced duration) on three hand-selected features, with no human listening test, no content-preservation metric (WER is not reported despite the "preserving spoken content" claim), and no baseline steering method.
- Single model. Results cover only CosyVoice3 with a 0.5B LM backbone; transfer to larger or differently structured TTS LMs is untested (§Limitations).
- Circular auto-interp evaluation. The same Gemini model writes and scores labels; the authors call for human evaluation and a scorer-model ablation (§Limitations).
- Partial auto-interp sweep. Detection scores exist only for a subset of layers, while modality and reconstruction statistics cover the full sweep (§Limitations).
- Temporal resolution. Speech-token timestamps are at 25 Hz and cannot localise sub-token acoustic onsets (§Limitations).
- Negative sampling. Negatives come from other features, testing label specificity but not robustness to near-neighbour confounds (§Limitations).
- Probe-to-steering link. The probe-based feature selection procedure is described only briefly; how the three steered features were chosen from the probing results is not fully specified in the text (§4.3, Appendix G).
- Some VCTK accent buckets (Welsh, New Zealand) are single-speaker, so those probes partly conflate accent with speaker identity (Appendix G).
Wiki Connections
- Autoregressive Codec TTS — Analyses and steers the frozen autoregressive LM of CosyVoice3 that generates 25 Hz discrete speech tokens, mapping which layers carry speech-side features.
- Prosody Control — Steering one SAE feature (3024) controls speech rate, measured as voiced duration ranging from 2.75 s to 10.57 s around a 3.96 s baseline.
- CosyVoice 3 — The frozen TTS system whose LM residual stream is decomposed and steered; all results are specific to this model.
- CosyVoice — Cited as the CosyVoice family reference for the supervised-semantic-token LM design analysed here.
- Qwen2.5 — Supplies the Qwen2.5-0.5B backbone inside CosyVoice3 on whose layers the SAEs are trained.
- Emilia — Provides the roughly 250M tokens used to train the SAEs and collect feature evidence.
- EmoSteer-TTS — Cited as prior activation-steering work for controlling speech generation; this paper steers through an SAE latent space instead of adding residual vectors.
- Sparse Autoencoders for Interpretable Emotion Control in TTS — Related SAE work that steers a frozen IndexTTS2 for emotion; not cited or compared in this paper.