Transcribe a recording
Choose a file, a model and a language. The computation runs on this computer, and your recording is not sent anywhere.
How it works
- Your browser reads the file and extracts the sound.
- The chosen model downloads once from this site, is checked against its fingerprint, then stays on your computer.
- Transcription moves through the recording thirty seconds at a time; the text appears as it goes, and you can listen and correct meanwhile.
- You save the result as Word, text, SRT or WebVTT. It is the only file that moves, and you are the one moving it.
Good to know
- A recent computer with Chrome or Edge gives the best results. A graphics card makes the Balanced and Accurate models much faster.
- The whole file is read into memory: an hour of audio takes about 250 MB once decoded.
- Nothing is kept when the tab closes. The browser warns you if you leave without saving.
- Speakers are identified once transcription has finished, then renamed in one go. The transcript should be proofread before it is shared.