Searcharxiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.25292

Does Expert Supervision Improve Text-to-Speech? An Empirical Study of Preference Optimization

Minh Duc Vu·Junchuan Zhao·Ye Wang
View original source ↗Read source document ↗Export metadata ↓

Abstract

Preference optimization promises to improve text-to-speech (TTS) systems by learning which speech outputs are better. External expert speech seems a natural source of preferred responses, especially when student generations contain errors. Yet demonstration quality does not establish whether the student can benefit from that supervision, or whether better target fitting will produce better speech. We examine this gap across two TTS students using real and synthetic expert speech, comparing preference objectives and direct versus bridge supervision while tracking fixed-sequence likelihoods and generation metrics. With expert supervision held fixed, objectives differ markedly in their sensitivity to preference-score scaling and the resulting generation quality. Likelihood diagnostics characterize these responses, but cannot establish their benefit: expert likelihood can increase while intelligibility worsens, or decrease while it improves. Bridge supervision highlights a distinction between mitigating degradation and improving the student: it can recover performance lost under direct training without consistently surpassing the pretrained base. Route benefits vary across objectives, students, and expert sources; neither route consistently dominates. Together, these findings distinguish demonstration quality, target fitting, and generated-speech quality as complementary criteria for assessing expert supervision.

Explore related subjects

cs.SDeess.AS

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Minh Duc Vu, Junchuan Zhao, Ye Wang. 2026-10-18. Does Expert Supervision Improve Text-to-Speech? An Empirical Study of Preference Optimization. https://arxiv.org/abs/2610.25292

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Deep learning-based multi-level airway collapse site localization on obstructive sleep apnea patients based on snoring sounds

Objective This study evaluated deep learning models, specifically ResNet-50 and Audio Spectrogram Transformer (AST), to identify multi-level upper airway collapses in obstructive sleep apnea (OSA) patients using snoring sounds. Methods Snoring recordings from 37 subjects undergoing drug-induced sleep endoscopy (DISE) between 2020 and 2021 were used to train models for two multi-label tasks: identifying obstructions at the Velum (V), Oropharynx (O), Tongue-base (T), and Epiglottis (E) levels, and the retropalatal (RP) vs. retroglossal (RG) levels. Results The AST model significantly outperformed ResNet-50 (P <0.05 after Bonferroni correction), effectively identifying V-, O-, and RP-level obstructions (F1-score: 0.59-0.87, Matthews Correlation Coefficient: 0.50-0.80, and AUC: 0.87-0.97). However, the performance of both models at the T, E, and RG levels remained limited due to small sample sizes, even with the Synthetic Minority Over-sampling Technique. A full-night recording analysis further demonstrated the feasibility of acoustic profiling for airway dynamics. Conclusion The AST models can effectively identify airway obstruction at V, O, and RP levels via snoring sounds. The proposed models offer a promising non-invasive tool for tracking overnight airway dynamics and assisting in clinical treatment planning.

cs.SD↗
arXiv2026

Speaker-Disentangled Remote Speech Detection of Asthma and COPD Exacerbations

Early detection of exacerbations in asthma and chronic obstructive pulmonary disease (COPD) is important for timely intervention. Speech has emerged as a promising tool for continuous, non-invasive respiratory disease monitoring. However, speech signals inherently carry speaker-identifiable attributes that may dominate model predictions, which may compromise both diagnostic accuracy and patient privacy. We propose an adversarial learning architecture that disentangles pathology-related acoustic patterns from speaker-identifiable attributes through gradient reversal-based training. To enhance clinical interpretability, we employ SHapley Additive exPlanations (SHAP) to quantify the contributions of acoustic features to pathology-related predictions versus speaker identity. On the TACTICAS dataset, the adversarial model significantly reduces speaker separability, while improving or maintaining classification performance. Specifically, it significantly improves the AUC for exacerbation type classification (from 0.65 to 0.73), enhances imbalance-sensitive metrics for respiratory status screening (PR-AUC and macro F1), and maintains cross-speaker generalization under the subject-wise split. Concurrently, the J-ratio decreases, confirming effective suppression of speaker information. This study demonstrates that adversarial disentanglement of speaker identity not only enhances patient privacy but also maintains or improves diagnostic accuracy for respiratory disease monitoring. Our framework offers a foundation for deployable, interpretable, and privacy-aware speech-based methods in chronic respiratory disease management.

cs.SD↗
arXiv2026

TTS-Q: Speaker-Aware Post-Training Quantization of Text-to-Speech Diffusion Transformers

Zero-shot text-to-speech (TTS) can clone a voice from seconds of audio. Yet the quality comes from billion-parameter diffusion transformers whose size and multi-step sampling put them out of reach of the on-device, assistive, and real-time applications that need them most. Post-training quantization (PTQ) shrinks such models without retraining, but it has been built for language and vision, not speech. At 4 bits, a strong quantizer leaves intelligibility close to full precision but consistently erodes speaker timbre, so error rates stay low while the voice drifts from the target speaker. We ask why quantization hurts speech, where the damage lives, and whether it can be repaired without retraining. To find out, we run the first 4-bit weight-and-activation (W4A4) PTQ study of a TTS diffusion transformer at 1B and 3.5B. The timbre cost comes from both the weight and the activation quantization, but only the activation part has the timestep structure a precision schedule can exploit. We therefore introduce TTS-Q, which selects hard-to-quantize items from task-matched calibration candidates and applies a budget-balanced mixed-precision schedule that assigns higher precision to late-step feed-forward activations. TTS-Q recovers timbre best at both scales, at no cost to intelligibility. Separately, the deployed calibrated-W4A4 reference reduces peak memory on GPU, providing a practical reference for resource-constrained TTS deployment.

cs.SD↗