People building private, local Thai voices now have a phonemization path designed for the language. The existing espeak-ng route did not segment unspaced Thai, placed leading vowels in writing order and omitted tones; measured against the TSync2 corpus, about 11% of its returned tokens contained no vowel.
The new Piper training path uses dictionary-based segmentation and syllabified pronunciation, with an explicit digit for each of Thai’s five tones. It also handles mixed text, numbers and alternate Unicode composition that could otherwise lose an entire utterance.
Across all 1,823 TSync2 utterances, the submitted verification produced no empty results and 34,405 syllables using symbols already present in Piper’s default phoneme map. The contribution also reports an intelligible single-speaker test voice. This is a foundation for training Thai voices, not a finished Thai voice release.