Pocket TTS ONNX (in-browser)
INT8 ONNX exports of Kyutai's Pocket TTS small English models, for running in the browser with onnxruntime-web, as used by pocket.ttstool.com.
| folder | model |
|---|---|
english_18m_2026-10/ |
english_18m_2026-10, 18.2M parameters |
english_18m_refl_2026-10/ |
english_18m_refl_2026-10, the same after reward fine-tuning |
english_7m_refl_2026-10/ |
english_7m_refl_2026-10, 6.9M parameters, reward fine-tuned |
english_2m_refl_2026-10/ |
english_2m_refl_2026-10, 1.9M parameters, reward fine-tuned |
Each folder holds bundle.json (metadata and the streaming state layout), tokenizer.model,
bos_before_voice.npy, and the text_conditioner, flow_lm_main, flow_lm_flow,
mimi_decoder and mimi_encoder models. Voices are cloned from reference clips, e.g. from
kyutai/tts-voices.
Exported with the workflow in
ken107/piper-browser-extension export/,
adapted from lomotron/pocket-tts-onnx-export.
The inference code is in web/src/pocket.ts.
License and attribution
Pocket TTS is developed by Kyutai and released under CC BY 4.0. These files are converted (ONNX, INT8 quantized) from Kyutai's weights and are distributed under the same license.
Prohibited use
As stated by Kyutai: use of the model must comply with all applicable laws and regulations and must not result in, involve, or facilitate any illegal, harmful, deceptive, fraudulent, or unauthorized activity. Prohibited uses include, without limitation, voice impersonation or cloning without explicit and lawful consent; misinformation, disinformation, or deception (including fake news, fraudulent calls, or presenting generated content as genuine recordings of real people or events); and the generation of unlawful, harmful, libelous, abusive, harassing, discriminatory, hateful, or privacy-invasive content. See Kyutai's model card.
Model tree for ken107/pocket-tts-onnx
Base model
kyutai/pocket-tts