---
license: cc-by-nc-4.0
library_name: onnx
tags:
  - text-to-speech
  - onnx
  - onnxruntime-web
  - hebrew
  - english
language:
  - en
  - he
---

# pocket-tts-onnx

[Kyutai's Pocket TTS](https://github.com/kyutai-labs/pocket-tts) exported to ONNX for
[pocket-tts-onnx](https://github.com/thewh1teagle/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](https://huggingface.co/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 |

```python
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`.