Transcribe speech in real time, on your device.
Press the button and start talking. The model downloads once, then nothing leaves this browser: no API key, no account, no server. Each finished phrase is tagged with how long it actually took.
Loading…
The badge on each line is the real measured time between you finishing a phrase and the final transcript arriving.
No microphone? Other ways to try it
Browsers keep their own default input, which can disagree with the one your operating system is using, and when it does it returns silence rather than an error. Naming the device here fixes that, and the choice is remembered. The microphone diagnostic shows what each layer is actually producing if this is still not working.
Bigger models are more accurate and take longer to download. All of them
still run comfortably in real time on a laptop. Word error rates are
LibriSpeech test-clean, measured on these exact quantized
models.
How we measure accuracy.
Live transcription, in your language
Three calls and two callbacks, whichever one you pick. The JavaScript tab is the code running on this page.