Building and Evaluating Fixed-Voice Thai TTS from Synthetic Speech
Kunat Pipatanakul, Potsawee Manakul, Warit Sirichotedumrong, Sittipong Sripaisarnmongkol, Pakorn Nathong, Phatrasek Jirabovonvisut
Wayu Research · Paxa Labs · Typhoon
Published September 3, 2026 · on the feed September 14, 2026 · ▲ 1 · arXiv 2609.03502
The abstract
In low-resource settings, deploying TTS typically requires choosing between a large voice-cloning model with costly inference or a compact fixed-voice system that requires a speaker-specific corpus. We study a third route: using a large voice-cloning model as a programmable data source to turn a short voice reference (e.g., 15 seconds) into a compact fixed-voice student trained entirely on synthetic speech. This setting makes pipeline design consequential: teacher errors become training targets, while filtering failed generations can reduce coverage of difficult texts. Thai further introduces challenges from ambiguous word boundaries, lexical tone, names and loanwords, numeric verbalization, and Thai-English code-switching. We study how text preparation, synthetic generation, quality filtering, rejection sampling, and frontend choices affect the resulting student, and where teacher limitations remain. We evaluate CER, Challenge-Set Keyword Accuracy, Prosody Pause Accuracy, speaker similarity, and speaking rate. The resulting 82M-parameter model, Wayu-Paxa-TTS-Edge, enables on-device Thai TTS without reference audio. It achieves 68.2% Challenge-Set Keyword Accuracy (85.5% of Gemini 3.1) and 91.4% pause precision, outperforming its OmniVoice teacher (89.9%) and reaching 94.8% of Gemini 3.1. It also achieves the lowest pause-placement error and intra-word pause rates among the three systems, and 3.7% and 1.1% CER on Thai and English, respectively. We open-source the model and evaluation framework for Thai TTS development.
The Brief
2:33A short podcast summary of the paper. Written and voiced by AI.
Transcript
Building and Evaluating Fixed-Voice Thai TTS from Synthetic Speech. By Kunat Pipatanakul and five others, from Wayu Research, Paxa Labs, and one other institution. Published on the third of September. No code repository is linked to it.
In low-resource settings, deploying text-to-speech usually forces a choice between two options. One is a large voice-cloning model, with costly inference. The other is a compact fixed-voice system, which needs a corpus recorded from one specific speaker. Thai adds its own difficulties on top. Word boundaries are ambiguous, the language has lexical tone, names and loanwords need handling, numbers have to be verbalized, and speakers switch between Thai and English mid-sentence.
The study takes a third route. The large voice-cloning model is used not as the deployed system but as a programmable data source. A short voice reference, around fifteen seconds for example, is turned into a compact fixed-voice student that is trained entirely on synthetic speech. That setup makes the pipeline consequential in a particular way. Whatever errors the teacher makes become training targets for the student. Filtering out the failed generations avoids that, but can strip out coverage of exactly the difficult texts the student most needs to learn.
The work therefore goes stage by stage and asks how each choice shapes the student: text preparation, synthetic generation, quality filtering, rejection sampling, and the frontend. It also asks where the teacher's limitations remain visible after all of that. Evaluation covers character error rate, Challenge-Set Keyword Accuracy, Prosody Pause Accuracy, speaker similarity, and speaking rate.
The resulting model, Wayu-Paxa-TTS-Edge, has eighty-two million parameters and runs Thai speech synthesis on device with no reference audio. It reaches sixty-eight point two percent Challenge-Set Keyword Accuracy, which is eighty-five point five percent of what Gemini three point one achieves. Its pause precision is ninety-one point four percent. That beats its OmniVoice teacher, at eighty-nine point nine, and reaches ninety-four point eight percent of Gemini's figure.
Among the three systems compared, it also has the lowest pause-placement error and the lowest rate of pauses falling inside words. Character error rate is three point seven percent on Thai and one point one percent on English.
The model and the evaluation framework are both open-sourced for Thai speech development. The broader lesson concerns the teacher. A big cloning model can serve as a source of data rather than a thing to deploy, provided the pipeline around it is designed knowing that the teacher's mistakes travel downstream.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
One voice, fifteen seconds of it, and a phone to run on. Something has to give.
You need a Thai speaking voice for an app that runs on the phone itself. All you have of the target voice is a short clip, about fifteen seconds long.
The road ahead
- 1 · The problem ~7 min You want a cheap Thai voice that runs on a phone, but you only have a short clip of that voice.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Use the big cloning model only to manufacture training speech, then train a small fixed-voice model on that speech alone.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Prepare Thai text, have the teacher speak it, filter bad takes, resample rejected texts, then train the small student.
- 6 · The experiment ~6 min You’ll guess: On pause precision in Thai, how did the 82M-parameter student compare with its OmniVoice teacher?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.