Transcribe audio on your device

How on-device transcription works

• The model comes to your audio. Your browser downloads a Whisper speech model from Hugging Face the first time, keeps it cached, and runs it on your own processor. The recording itself is never sent anywhere.
• Your device picks the model. With WebGPU, the page uses Whisper base for better accuracy. Without it, it uses Whisper tiny, a small English-only model that stays quick enough for a phone.
• Long recordings are read in pieces. The audio is resampled to 16 kHz mono and transcribed in 30-second windows that overlap, so words that fall on a boundary are not lost.
• Speakers are optional. Tick Identify speakers and a second on-device model groups the transcript by voice. Telling it how many people speak, up to 20, gives cleaner labels than auto-detect.

Good to know

What about languages other than English?

On a device without WebGPU the model only knows English, so other languages come out poorly. For those, use Audio to text, which runs a larger Whisper model on our own workers and detects the language for you.

Why is it slow on my laptop?

Without WebGPU the model runs on the processor alone, and speed depends on how fast it is. Close other heavy tabs, try a recent Chrome or Edge, or split a long recording into shorter files.

Can I keep the transcript?

Yes. Copy it, or download it as a .txt file. Nothing is saved on our side, so save it before you close the page.