Building and Evaluating Fixed-Voice Thai TTS from Synthetic Speech
Kunat Pipatanakul, Potsawee Manakul, Warit Sirichotedumrong, Sittipong Sripaisarnmongkol, Pakorn Nathong, Phatrasek Jirabovonvisut
cs.CL · cs.AI
Quick verdict
Distills a 15s reference into an 82M fixed-voice Thai TTS trained on synthetic speech; rigorous ablations, but no perceptual test or significance stats.
Summary
The paper presents Wayu-Paxa-TTS-Edge, an 82M-parameter fixed-voice Thai-English TTS model trained entirely on synthetic speech generated by the 600M-parameter zero-shot voice-cloning teacher OmniVoice from a 15-second reference. The pipeline comprises Thai text preparation with LLM verbalization of digits and code-switched English, teacher sampling over 12 frozen voice references, quality filtering (CTC-based pronunciation checks on hard tokens plus pause, speaking-rate and duration controls; 23% overall rejection), and rejection sampling that re-renders failures to preserve difficult-text coverage. A Kokoro/StyleTTS2 student with a new script-routed Thai-English phoneme frontend is trained on 54.35 hours of filtered bilingual data. Evaluation combines CER, a 1,531-item Challenge-Set Keyword Accuracy across five error-prone categories, pause-placement metrics on 210 long sentences, speaker similarity, and speaking rate. The final model reaches 68.2% keyword accuracy (teacher 72.8%, Gemini 3.1 Flash 79.8%), 3.7% Thai and 1.1% English CER, and 91.4% pause precision, surpassing its teacher (89.9%) with the lowest PPER (6.7%) and intra-word pause rate of the three systems. Oracle best-of-K sampling (K=118) exposes 15.1pp of recoverable teacher headroom, and keyword frequency in teacher training data correlates with sampling difficulty (rho=-0.437). Main limitations: no perceptual quality evaluation, no statistical significance testing, small evaluation sets, and evaluation instruments overlapping the training filter.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A 15-second voice reference can be converted into an 82M fixed-voice Thai TTS trained entirely on synthetic speech from a zero-shot cloning teacher, deployable on-device without reference audio.
- Pipeline design — especially rejection sampling of filtered failures — matters more than corpus size: resampling rejected candidates reverses the keyword-accuracy loss caused by strict filtering at scale (Table 5).
- The student surpasses its teacher on pause placement (91.4% vs 89.9% precision; PPER 6.7%, intra-word 1.4%) while trailing on Challenge-Set keyword accuracy (68.2% vs 72.8%) and Gemini 3.1 (79.8%).
- Oracle best-of-K teacher sampling (87.9% at K=118 vs 72.8% at K=1) shows 15.1pp recoverable headroom, with residual failures concentrated on keywords underrepresented in the teacher's training corpus (ρ=-0.437).
- A 15-second Isan-accented reference transfers voice identity and dialect forms to the fixed-voice student (Isan CER 5.5% vs teacher 6.6%; similarity 0.842 vs 0.854) at the cost of Central-Thai performance (CER 3.4%→5.1%).
Strengths
Multi-dimensional evaluation designed for Thai-specific failure modes
Section 3 builds a 1,531-item Challenge Set across five categories (code-switching 391, names 310, rare words 410, informal spelling 210, long sentences 210) plus pause metrics on 210 sentences; Table 3 shows keyword accuracy separates Gemini 3.1 (79.8%) from OmniVoice (72.8%) where CER (3.3% vs 4.6%) understates the gap.
Carefully instrumented and validated pause metric
Appendix A.2 validates three forced aligners against duration-predictor boundaries on ~1,500 clips (wav2vec2 r=0.984 per-speaker; MMS_FA rejected at r=0.724), publishes all fixed scorer constants (Table 13), discusses residual error sources, and controls PPER gaming via under-pausing by reporting pauses/clip.
Controlled ablations exposing the quality-coverage tradeoff
Honest teacher-limit analysis quantifying the distillation ceiling
Transparency and released artifacts
Weaknesses
No perceptual quality evaluation
No statistical significance testing or variance reporting
Potential evaluator-pipeline circularity
Missing baseline for the central deployment argument
Abstract framing selectively favorable
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This method-plus-empirical paper from Wayu Research, Paxa Labs, and Typhoon builds **Wayu-Paxa-TTS-Edge**, an 82M-parameter fixed-voice Thai-English TTS system, by distilling the 600M-parameter zero-shot voice-cloning model OmniVoice (Zhu et al., 2026) into a Kokoro/StyleTTS2 student trained **entirely on synthetic speech** derived from a 15-second voice reference. Claimed contributions: (1) an end-to-end recipe (text preparation → teacher sampling → quality filtering → rejection sampling → student training); (2) a Thai-specific evaluation framework separating CER, Challenge-Set Keyword Accuracy (1,531 items), Prosody Pause Accuracy (210 long sentences), speaker similarity, and speaking rate; (3) controlled ablations of pause filtering, rejection sampling, pretrained initialization, and frontend policy; (4) a teacher-limit analysis via oracle best-of-K sampling plus an Isan adaptation study. The paper is notably candid in places: it states the student trails its teacher on keyword accuracy, flags Gemini's parameter count as a proxy, and devotes Section 5 to what distillation cannot fix.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Wayu-Paxa-TTS-Edge?
An open-source 82M-parameter Thai-English text-to-speech model with a fixed voice that runs on-device without reference audio. It was trained entirely on synthetic speech generated by a larger voice-cloning teacher model (OmniVoice, 600M parameters) from a 15-second voice sample. It beats that teacher on pause placement (91.4% vs 89.9% precision) but scores slightly lower on difficult-word pronunciation accuracy (68.2% vs 72.8%).
How does an 82M TTS model compare with Gemini's TTS?
On this paper's Thai benchmarks, Wayu-Paxa-TTS-Edge reaches 68.2% challenge-set keyword accuracy versus 79.8% for Gemini 3.1 Flash TTS, and 91.4% pause precision versus 96.4%; Thai character error rate is 3.7% versus 3.3%. Gemini also has fewer misplaced-pause clips despite producing about five times more pauses per clip. The student's practical advantages are open weights, on-device deployment, and no reference audio needed at inference.
Why train TTS on synthetic speech instead of real recordings?
In low-resource languages like Thai, recording and licensing a single-speaker corpus is expensive. This paper's route uses a large zero-shot voice-cloning model as a programmable data source: any text can be rendered in the target voice, then filtered for pronunciation and pause errors before training a compact student model. The tradeoff is that teacher errors become training targets, and rare expressions the teacher cannot pronounce remain unresolved (66% coverage for unseen keywords even at 118 samples).
Is the model and evaluation code available?
Yes. The paper states that the model (Wayu-Paxa-TTS-Edge) and the evaluation framework (Thai-TTS-Eval, github.com/wayu-research/thai-tts-eval) are open-sourced, with teacher (k2-fsa/OmniVoice) and student (hexgrad/Kokoro-82M backbone) components on Hugging Face. Regenerating the training corpus additionally requires API models (DeepSeek verbalization, Gemini pause masks), so third parties can most easily verify evaluation numbers rather than rebuild the corpus.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Thai script embedded inside figure graphics (e.g., the Figure 5 waveform-panel labels) is garbled by PDF extraction; all prose, tables, and footnotes are intact and legible.
- The provided text is complete through Appendix A.2 and Table 14, ending with a full sentence — no truncation was observed; the arXiv abstract snippet in the metadata is cut mid-sentence but the full abstract appears in the paper body.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.