pocket-tts-gguf

The GGUF Q8_0 weights, one per language, that tts-web's Pocket TTS engine loads in the browser.

kyutai/pocket-tts-without-voice-cloning, by Kyutai

Q8_0 GGUF quantization for the browser by Idle Intelligence.

CC-BY-4.0

Q8_0 GGUF quantization of Kyutai’s Pocket TTS in six languages (English, French, German, Spanish, Portuguese, Italian), one file per language under languages// as in Kyutai's repo. Decoder path only (the Mimi encoder is excluded, TTS never needs it): transformer backbone, flow matching network, Mimi decoder and decoder transformer, shrunk from 236MB to 128MB for English. Runs via a tiled WASM SIMD128 quantized matmul kernel for about 2x realtime on desktop Chrome.

Type
model
Runs
browser, WASM
Size
~97M params, 128-134MB per language (Q8_0, decoder path only)
Demo (GitHub Pages)
https://idle-intelligence.github.io/tts-web/web/
Source
idle-intelligence/tts-web
HuggingFace
idle-intelligence/pocket-tts-gguf
Status
maintained

Used on: tts

← Home