How it works
After your transcript is ready, click Detect speakers in the side panel. A small open speaker model (pyannote segmentation 3.0, about 6 MB) downloads once and listens for changes in voice. It works through the recording in overlapping 30 second windows and matches voices across windows, so the same person keeps the same label from start to finish. Each line of the transcript then gets the speaker with the most speech in it.
Getting the best labels
- Two to four clear voices give the most reliable results. Large group calls and heavy crosstalk may need a few manual fixes.
- Rename speakers once in the side panel. The names appear in Word, PDF, subtitles and every other export.
- Fix any single line by clicking its label.
- Use a recording where each person is reasonably close to the microphone. Echo and music make voices harder to tell apart.
Why on-device matters here
Speaker identification usually means uploading the whole recording to a cloud service, because it needs to hear every voice. Here it runs in your browser like the transcription itself, which makes it suitable for confidential interviews, HR conversations and research with participants.
What happens to your recording
Your browser reads the file from your device and runs the speech model locally. The audio, the video and the transcript are never sent to a server, so there is nothing for us, or anyone else, to store, train on or lose in a breach. Finished transcripts are saved only in your browser’s storage, where you can export or delete them in your library. The only thing that comes from the internet is the speech model itself, once. You can check this by turning off Wi-Fi after the first transcript: it keeps working.