stt-web

Transcribes speech to text live in the browser, in English and French.

Browser-native speech-to-text, 100% client-side via Rust/WASM + WebGPU. Microphone audio runs through a Mimi codec (CPU) and a Q4-quantized 16-layer STT transformer (WebGPU) end to end in Chrome/Edge. Weights are quantized from Kyutai’s stt-1b-en_fr; not affiliated with Kyutai Labs.

Runs
browser, WASM, WebGPU
Model
kyutai/stt-1b-en_fr, ~989M params, model licence CC-BY-4.0; code licence MIT
Demo (GitHub Pages)
https://idle-intelligence.github.io/stt-web/web/
Source
idle-intelligence/stt-web
HuggingFace
idle-intelligence/stt-1b-en_fr-q4_0-webgpu
Perf
~70ms/frame steady state (Mimi ~21ms + STT ~50ms), RTF 0.91-0.95x, Apple M2 (10-core GPU)
Status
maintained

Used on: stt, stt + llm + tts

Blog post: Claude and stt-web

← Home