8000 kiteva/models at main · leminh0796/kiteva · GitHub
[go: up one dir, main page]

Skip to content

Latest commit

 

History

History
 
 

models

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 

README.md

Local speech models

Downloaded weights stay in visible, engine-specific directories and are ignored by Git (/models/kokoro/ and /models/pocket-tts/ in .gitignore). A fresh clone has none of them. Each engine's setup command downloads its files once. Synthesis never downloads models.

Kokoro

  • kokoro/kokoro-v1.0.onnx: Kokoro model.
  • kokoro/voices-v1.0.bin: Kokoro voice data.

Run npm run speech:setup to install Kokoro. Existing HyperFrames cached files are copied when available; otherwise setup downloads them from the kokoro-onnx model release. Original cache files are preserved. The Python environment lives separately in .data/speech-venv.

Pocket TTS

384E Pocket TTS is implemented as its own provider (src/server/audio/providers/pocket-tts.ts), installed as pocket-tts>=3.3.0,<4 (scripts/pocket-tts-requirements.txt).

  • pocket-tts/languages/<language>/: a local-only config.yaml, a ready.json readiness manifest and voices/ with the exported preset voice states, for each of the seven languages in src/shared/pocket-tts-catalog.json (Dutch, English, French, German, Italian, Portuguese and Spanish).
  • pocket-tts/hub/, pocket-tts/huggingface/ (and pocket-tts/torch/ if created): the Hugging Face and PyTorch caches the setup script points at, holding the downloaded model weights (kyutai/pocket-tts-without-voice-cloning). Synthesis loads them with Hugging Face offline mode on.
  • The current setup script doesn't write config.yaml, ready.json or voices/ directly under pocket-tts/. If your copy has them at that level, they aren't read by the per-language readiness check.

Run npm run speech:setup:pocket to install every language, or npm run speech:setup:pocket -- --language italian for one. The Python environment lives separately in .data/pocket-tts-venv. POCKET_TTS_PYTHON can point at another environment with a supported package version. Voice cloning isn't exposed.

Shared

Python environments and generated project audio stay separate from these model files. Model weights and voice data keep their upstream licenses.

Provider TypeScript implementations and their shared interface are tracked under src/server/audio/providers. See speech setup.

0