Project / local speech

Offline speech to text, in real time.

A small multilingual transcription workspace powered by Whisper in WebAssembly. Your microphone audio is processed in the browser, not sent to an application server.

Live transcription

Preparing local runtime

Loading the browser-side speech engine.

Transcript

Words will appear here once listening begins.

Model

tiny · Q5_1 · multilingual
31 MB download · kept in this browser

Spoken language Auto-detect locks after consistent local results.

First use The model is downloaded directly to this browser.

Private by design

The page asks for microphone access only when you start listening. Audio is captured, resampled, and decoded locally in the active tab.

Made for live speech

It keeps a short rolling audio context and sends contexts to Whisper in the background. Overlapping hypotheses are reconciled before text is settled, so the transcript can evolve without obvious repeats.

Choose the trade-off

Tiny starts faster and uses less memory. Base is larger, but can be the better choice when recognition quality matters more than initial load time.