Original title: Building and Evaluating Fixed-Voice Thai TTS from Synthetic Speech
Kunat Pipatanakul, Potsawee Manakul, Warit Sirichotedumrong, Sittipong Sripaisarnmongkol +2 · September 2026
Building text-to-speech models for less common languages usually forces a hard choice. Teams must run a massive, slow computer program or spend months recording one person's voice. Researchers tested a third option for the Thai language. Thai is tricky because it has complex tones, blended foreign words, and no spaces between words. To solve this, the team took a 15 seconds voice sample and used a large voice-cloning teacher model to create fake training audio. Then, they trained a compact student model solely on this computer-made speech. Their small 82M-parameter model, called Wayu-Paxa-TTSText-to-Speech technology that converts written text into spoken voice.-Edge, runs directly on regular devices without needing reference audio. It placed spoken pauses more accurately than its own teacher and made fewer awkward mid-word breaks than larger systems, showing that compact voices can learn effectively from synthetic speechComputer-generated audio that sounds like human speech. alone.
You want to learn to cook a family recipe after hearing your grandmother describe it once. Normally, you'd need her to guide you through many attempts, but she's not available. So, you imagine how she would say the steps and practice those imagined instructions. In this paper, a large AI system does the imagining to create many speech examples, and a small system learns to speak by practicing those, so it can eventually talk like the grandmother without her constant presence.
For languages like Thai with limited recorded data, this method allows building voice systems from just a short audio sample. It means voice assistants and accessibility tools can run on phones without needing extensive speaker recordings, lowering barriers for local language support.
The method depends on the quality of the larger teacher model, which may have errors that become training data. It is evaluated specifically for Thai, so its effectiveness for other languages is not established. The model does not fully match the performance of larger state-of-the-art systems like Gemini 3.1.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Building and Evaluating Fixed-Voice Thai TTS from Synthetic Speech — Kunat Pipatanakul, Potsawee Manakul, Warit Sirichotedumrong et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.