---
license: cc-by-4.0
language:
- en
- fr
- de
- it
- pt
- es
base_model:
- kyutai/pocket-tts
pipeline_tag: text-to-speech
library_name: phoonnx
tags:
- tts
- onnx
- onnxruntime
- pocket-tts
- kyutai
- phoonnx
---

# Pocket TTS — ONNX mirror for phoonnx

ONNX weights for [Kyutai Pocket TTS](https://kyutai.org/tts/), packaged for the
[phoonnx](https://github.com/TigreGotico/phoonnx) `pockettts` engine.

Pocket TTS is a 100-million-parameter text-to-speech model. It runs faster than real time
on two CPU cores. This mirror holds one weight bundle per language, in full precision and
in 8-bit, together with the published speaker states.

## Languages

| Directory | Language | Transformer layers |
|---|---|---|
| `english_2026-04` | English | 6 |
| `french_24l` | French | 24 |
| `german` | German | 6 |
| `italian` | Italian | 6 |
| `portuguese` | Portuguese | 6 |
| `spanish` | Spanish | 6 |

## Contents of a bundle

| File | Purpose |
|---|---|
| `bundle.json` | Sample rate, latent size, chunk limit and the two state manifests |
| `tokenizer.model` | SentencePiece tokenizer, 4000 units |
| `bos_before_voice.npy` | Embedding prepended to a cloned voice |
| `text_conditioner.onnx` | Token ids to text embeddings |
| `flow_lm_main.onnx` | Autoregressive transformer with explicit state |
| `flow_lm_flow.onnx` | Flow network for one latent frame |
| `mimi_decoder.onnx` | Latent frames to 24 kHz audio |
| `mimi_encoder.onnx` | Reference audio to latent frames, for voice cloning |
| `voices/<name>.safetensors` | Published speaker states, 26 per language |

Every graph except the encoders also ships as `*_int8.onnx`.

## Graph interfaces

```
text_conditioner : token_ids                          -> embeddings
flow_lm_main     : sequence, text_embeddings, state_* -> conditioning, eos_logit, out_state_*
flow_lm_flow     : c, s, t, x                         -> flow_dir
mimi_decoder     : latent, state_*                    -> audio_frame, out_state_*
mimi_encoder     : audio                              -> latents
```

Audio is 24 kHz mono. The model produces 12.5 latent frames per second, and each frame
decodes to 1920 samples. The `state_*` tensors are described by the state manifests in
`bundle.json`: each entry gives the tensor name, shape, dtype and how to fill it at the
start of a stream.

A voice is a *state*, not an embedding vector: it is the transformer state after the model
has consumed a speaker's audio. The manifests map a `voices/*.safetensors` file onto the
state tensors of `flow_lm_main`.

## Usage

```python
from phoonnx.model_manager import TTSModelManager

manager = TTSModelManager()
voice = manager.download_voice_by_id("pockettts/en/alba").load()
audio = voice.synthesize_wav("Hello world.")
```

## Attribution

The model is the work of [Kyutai](https://kyutai.org/), released as
[`kyutai/pocket-tts`](https://huggingface.co/kyutai/pocket-tts). The ONNX graphs were
exported by the community and are mirrored from
[`KevinAHM/pocket-tts-onnx`](https://huggingface.co/KevinAHM/pocket-tts-onnx). The speaker
states come from
[`kyutai/pocket-tts-without-voice-cloning`](https://huggingface.co/kyutai/pocket-tts-without-voice-cloning).

This repository adds no new weights. It exists so that phoonnx can fetch a stable set of
files from one place.

## License

The weights keep their upstream license, Creative Commons Attribution 4.0 International
(CC BY 4.0). The full text is in `LICENSE`. Give credit to Kyutai when you use them.

Kyutai asks that the model is not used for voice impersonation without consent, for
deception, or for unlawful or abusive content.