pocket-tts-onnx
Kyutai's Pocket TTS exported to ONNX for pocket-tts-onnx, including a Hebrew IPA adapter. Runs in the browser on onnxruntime-web, or in Python with onnxruntime and numpy.
Try it: https://thewh1teagle.github.io/pocket-tts-onnx/
Web asset set
The browser app fetches these; they are hosted here rather than in a GitHub release because release assets are not served with CORS headers.
| file | size | what |
|---|---|---|
model.onnx |
177 MB | the streaming step graph, int8, adapter bundled |
encoder.onnx |
39 MB | the mimi voice encoder, fetched only when cloning |
assets.json |
1.4 MB | config, tokenizer and voice conditionings |
manifest.json |
tiny | what to fetch and how big it is |
renikud.onnx |
21 MB | Hebrew G2P, mirrored from thewh1teagle/renikud |
Single-file models
For the Python package, where everything rides inside one file:
| file | size |
|---|---|
pocket-tts-english.onnx |
228 MB |
pocket-tts-english-ipa.onnx |
231 MB |
from pocket_tts_onnx import PocketTTS
tts = PocketTTS("pocket-tts-english-ipa.onnx")
samples, sample_rate = tts.create("Hello world.", voice="alba")
Voices baked in: alba, cosette, javert, male1.