Kokoro Thai — th_v3p (serving bundle)
Serving bundle for the Thai Kokoro student model: 12 speakers, 24 kHz output.
Contains the inference-only checkpoint, the speaker roster with its reference
recordings, prebuilt length-aware voice packs, and the kokoro_thai frontend
code the runtime imports.
Consumed by kokoro-tts-web — set KOKORO_HF_REPO to this
repo id and the backend downloads and uses it as its model root.
Layout
| Path | What |
|---|---|
onnx/ | the serving path: ONNX graphs, per-speaker style vectors, numpy voice packs. Runs on onnxruntime alone — no torch, no checkpoint |
runs/th_v3p/kokoro_thai_ep8.pth | inference weights (7 modules, no optimizer/discriminator state), for the torch backend |
kokoro_thai/, train.py, assets/ | G2P + model code imported at runtime |
data/seeds/ | roster.json and one reference wav per speaker |
voicepacks/ | prebuilt per-speaker length-aware style packs (pace="steady") |
samples/ | one pre-rendered preview per voice, so a deployment never synthesizes them at boot |
Text frontend
kokoro_thai/g2p.py is bilingual: Thai runs go through tltk, Latin runs through
misaki-en (Kokoro's own English G2P), both landing in the shared 114-token vocab.
A deployment therefore needs misaki[en] and a pinned en_core_web_sm alongside
tltk — see requirements-deploy.txt in kokoro-tts-web.
Derived from Kokoro-82M (StyleTTS2-lite), fine-tuned on Thai.