kairos-voice

Kyutai's Pocket TTS for the browser, French and English, mirrored so the kaairos video editor downloads files that cannot change under it. It reads a voiceover's words on the visitor's own machine, on onnxruntime-web 1.30 (WebAssembly, one thread), and nothing is sent anywhere.

Files

Two bundles in the volume-board/pocket-tts-onnx@1 format, one per language: the flow language model's backbone and flow head and the Mimi decoder (int8), the text embedding table (float16), the tokenizer, one predefined voice (the flow model's attention cache after hearing the speaker, float16) and a manifest of shapes and generation settings.

folder	from	voice
fr/	Vincweb/pocket-tts-french-onnx, byte for byte	Estelle
en/	exported from Kyutai's English weights with the same scripts (below)	Alba

The sha256 of every file is in the kaairos repository, training/voice/export.sha256.

What it is, and the terms

• Pocket TTS — Kyutai (kyutai-labs/pocket-tts, weights kyutai/pocket-tts-without-voice-cloning), CC BY 4.0. French weights at revision 1e08e6a23401048648a9fdcfde2f89348215c2a7, English at a7056f505a4a8ca72d8b5fb1440071d6f1a35739.
• Changes made: a format conversion, not a new model — the PyTorch modules exported to ONNX with their streaming state as graph inputs and outputs, then dynamic int8 quantisation of the MatMul and Gemm weights (onnxruntime.quantization.quantize_dynamic). French converted by Vincweb; English by the kaairos project with Vincweb's scripts, pocket-tts 4.0.1, torch 2.14.1 (CPU). The English export was checked against PyTorch on one flow step (max difference 1.5e-6, end-of-speech logit −10.7658 both).
• Voices: Estelle — kyutai/tts-voices, unmute-prod-website/developpeuse-3.wav, CC0. Alba — Alba Mackenna, kyutai/tts-voices, alba-mackenna/casual.wav, CC BY 4.0.
• Kyutai asks, and so does this mirror: no voice impersonation or cloning without explicit and lawful consent, no deceptive use, and no presenting generated speech as a genuine recording of a real person or event.