tts-web
Converts text to speech in the browser with two lightweight voice models; Pocket TTS speaks six languages.
Browser-native text-to-speech, 100% client-side via Rust/WASM. Two engines: Pocket TTS (autoregressive, Mimi codec decoder, streaming) and KittenTTS (single forward pass, StyleTTS2 distilled, 14M params). Weights on HuggingFace as GGUF/safetensors; 0.23-0.24 RTF native. Pocket TTS speaks English, French, German, Spanish, Portuguese and Italian, each language a Q8_0 GGUF quantization (~130MB) of Kyutai’s checkpoint, with voices loaded from Kyutai’s repo at the revision the weights came from. In F32 it matches the official implementation frame for frame.
Includes
KittenML/kitten-tts-nano-0.8, by KittenML
Safetensors conversion for candle inference by Idle Intelligence.
Apache-2.0
modelkyutai/pocket-tts-without-voice-cloning, by Kyutai
Q8_0 GGUF quantization for the browser by Idle Intelligence.
CC-BY-4.0
modelkyutai/pocket-tts-without-voice-cloning, by Kyutai
Channel-wise INT8 quantization by Idle Intelligence.
CC-BY-4.0
- Runs
- browser, WASM, native
- Demo (GitHub Pages)
- https://idle-intelligence.github.io/tts-web/web/
- Source
- idle-intelligence/tts-web
- HuggingFace
- idle-intelligence/pocket-tts-gguf
- Perf
- Pocket TTS: 2.28x realtime in WASM/Chrome (TTFB 0.41s), M-series Mac
- Status
- maintained
Learn more
Used on: tts, llm + tts, stt + llm + tts