Client-side speech recognition with asr4all

asr4all transcribes as you speak, entirely on your device: int8 models from 38 MB, running in WebAssembly on a single CPU core.

Try asr4all

asr4all-s38 MB · fastest asr4all-m72 MB asr4all-l123 MB · most accurate

The model downloads once and is then kept in your browser's cache. Audio never leaves the page.

Wakes on speech

With Skip silence on, a 2.5k-parameter VAD listens and the recognizer runs only while you talk.

A small second pass adds punctuation and capitals as the transcript grows.

Recognition runs locally, in WebAssembly. Your audio is never uploaded.