ICICC · 2025 · Conference
Parshav Singla et al. (Thapar Institute of Engineering and Technology / Samsung Research and Development Institute Bangalore) · → Paper · Demo: ? · Code: ?
Frames read-to-conversational speaking-style conversion as mel-spectrogram resynthesis with a HiFi-GAN vocoder (named PACC, Prosodic Adjustment with Conversational Context), and reports objective and MOS gains over an unspecified baseline.
Problem
Read speech tends to have flat intonation, uniform rhythm, and little emphatic stress, whereas conversational speech varies all three. Voice interfaces such as virtual assistants, customer-service agents, and language-learning tools that rely on read-style output can sound robotic as a result. The authors argue that earlier rule-based approaches, which adjust only pitch and duration, cannot reproduce the variability of natural conversation. They set out to convert read speech into conversational speech while keeping computational cost low enough for on-device use.
Method
The system is called Prosodic Adjustment with Conversational Context (PACC). The paper describes it as a deep-learning framework with a “context-aware module” that adjusts intonation, stress, and rhythm according to the conversational scenario. That module is never specified: there is no architecture, input description, objective, or diagram for it. The only concrete pipeline described is that input read-speech recordings are segmented, converted to mel-spectrograms by STFT and mel-scale mapping (the introduction calls them “MFCC spectrograms” in places), and passed to NVIDIA’s HiFi-GAN vocoder (HiFi-GAN) to produce waveforms. The HiFi-GAN description (a convolutional generator with multi-resolution discriminators that are discarded after training) is background on the existing vocoder, not a new design.
Training is described only by its hyperparameters: 20, 30, and 40 epochs, batch sizes of 16, 32, and 64, learning rate 5e-5, dropout 0.1, and the Adam optimizer (§7.1). The training data is described as curated high-quality recordings of read and conversational speech at 16 kHz or higher, with speaker and text metadata (§5). No corpus is named, no sizes or speaker counts are given, and the paper does not say how read and conversational utterances are paired. The abstract credits “additional training on speech dataset” and cites the LJ Speech Dataset at that point. The paper does not say whether HiFi-GAN is trained from scratch or fine-tuned from NVIDIA’s released checkpoint. Model size is not reported.
Key Results
Table 2 compares the “HiFi-GAN Model” with a “Baseline Model” whose identity is never stated. The proposed system has MCD 0.57 versus 4.12, PESQ 4.64 versus 3.47, and RMSE 0.29 versus 0.35. Lower is better for MCD and RMSE and higher is better for PESQ, so the proposed system leads the baseline on all three rows (§7.2). In the subjective evaluation, the proposed system averages MOS 4.2 against the baseline’s 3.6 (§7.3). The paper gives no listener count, sample count, confidence intervals, or test set.
Several reported numbers are inconsistent with the paper’s own definitions. PESQ is 4.64 in Table 2, but §6 gives the PESQ range as -0.5 to 4.5. Figure 2 plots “MOS” values between about 2 and 8 across NB/WB/SWB/FB bandwidth categories, but §2.4.1 defines MOS on a 1 to 5 scale. Figure 2 also contains several unlabeled data series and does not identify the systems or the x-axis units.
The introduction also claims a 30% increase in listener preference and a 20% reduction in perceived monotonicity relative to “traditional methods”. No experiment in §7 reports either figure. Pitch Contour Distortion is defined in §2.4 and §6, but no PCD value is reported. The remaining evidence is a mel-spectrogram and a waveform plot of one output (Figures 3 and 4), which are discussed qualitatively.
Novelty Assessment
The only component that the paper actually specifies is an off-the-shelf HiFi-GAN vocoder applied to mel-spectrograms of the input speech. This amounts to engineering integration at most. The claimed novelty, conversational-context prosody adjustment, has no described mechanism, so it cannot be evaluated or reproduced. The evaluation measures fidelity to a reference (MCD, PESQ, RMSE) and general naturalness (MOS). It does not measure how far the output’s prosody moves away from the read input or toward conversational speech, so it does not test whether a style conversion took place. With an unnamed baseline and unnamed test data, the reported gains cannot be placed relative to other work.
Field Significance
Low. The paper is a narrow application report on the read-to-conversational speaking-style conversion task. Its evidence for that task is weak because the prosody-modification component is unspecified and several reported metric values fall outside their stated ranges. It mainly serves as a cautionary data point: vocoder resynthesis quality metrics are not, on their own, evidence of style conversion.
Claims
- complicates: Fidelity metrics computed against a reference (MCD, PESQ, waveform RMSE) together with naturalness MOS do not establish that a speaking-style conversion took place; style conversion claims need measures of prosodic change relative to the source, such as pitch contour or rhythm statistics.
Evidence: The pipeline vocodes mel-spectrograms of the input read speech with HiFi-GAN and reports only MCD, PESQ, RMSE, and MOS. Pitch Contour Distortion is defined, but no value is reported, and no comparison against the read input is given. (§2.4, §7.2, Table 2, §7.3)
- complicates: Objective and subjective quality scores for synthesized speech need checks against each metric’s defined range and scale before they can be compared across papers.
Evidence: Table 2 reports PESQ 4.64, although §6 gives the PESQ range as -0.5 to 4.5. Figure 2 plots MOS values up to about 8, although §2.4.1 defines MOS on a 1 to 5 scale. (§6, §7.2, Table 2, §2.4.1, Figure 2)
- supports: Neural GAN vocoders driven by mel-spectrograms can produce output that listeners rate as natural, but naturalness ratings alone do not distinguish vocoder quality from conversion quality.
Evidence: The HiFi-GAN system receives MOS 4.2 against 3.6 for an unspecified baseline. Listener count, test material, and confidence intervals are not reported. (§7.3)
Limitations and Open Questions
Warning
The PACC context-aware prosody-adjustment module that the paper presents as its contribution has no architecture, inputs, or training objective. The baseline system and the training and evaluation data are unnamed. The reported results cannot be reproduced or attributed to any component beyond the off-the-shelf HiFi-GAN vocoder.
- PESQ (4.64) falls outside the paper’s own stated range, and Figure 2’s MOS values exceed the stated 1 to 5 scale, so the reported metrics are not reliable.
- The preference (+30%) and monotonicity (-20%) figures in the introduction have no supporting experiment.
- The paper does not explain how conversational targets are obtained for read inputs (parallel pairs, style labels, or another mechanism), so the source of the supervision is unclear.
- The paper presents efficiency and edge-device suitability as advantages but reports no latency, real-time factor, or model size.
Wiki Connections
- GAN Vocoder — The pipeline relies on a HiFi-GAN vocoder applied to mel-spectrograms as its only specified synthesis component.
- Voice Conversion — Read-to-conversational speaking-style conversion is a speech-in, speech-out style-transfer task, evaluated here with reference-based spectral and waveform distortion metrics.
- Subjective Evaluation — Reports listener MOS for the converted speech against a baseline but gives no details of the listening protocol, and its Figure 2 MOS values exceed the stated 1 to 5 scale.
- HiFi-GAN — Used directly as the synthesis backbone; the paper adds no architectural changes to it.