Speaker Separation for Audio

A transcript gives you text where lines are tagged “Speaker 1”, “Speaker 2”. Here the result is sound: a separate audio file for each person that you can level, clean up or cut in your editor independently of the others.

The tool finds those places and marks them with a strip above the timeline, but it does not separate the voices there: the overlapping audio goes into both people’s tracks. When people take turns, each line goes only into its speaker’s track.

Yes, if you choose “Full length”. Every track has the same length, sample rate and channels as the original, with silence where someone else is speaking. Place all tracks at zero on the timeline and together they give back the original recording.

Click the right number in the “Actual number of people” row — the split is recalculated at once, without processing again. If one person ended up in two rows, choose “Same person as…” in one of them and the rows become one track.

Voices are told apart by their timbre, so two similar voices sometimes end up in one track. In that case, click the right number in the “Actual number of people” row: when the number of people is given, the voices split noticeably more accurately. Results are best when people take turns and are recorded at roughly the same volume.

Any: voices are told apart by their sound, not by their words. You can even split a recording where people speak different languages.