Diarization
Labelling who spoke when in an audio recording, so a transcript is split into turns by speaker.
Why it matters. Needed for meetings, calls and interviews, where a transcript without speakers is hard to use.
How it works
Speaker diarization splits audio into segments and labels each one with a speaker, answering 'who spoke when'. It usually assigns generic labels such as Speaker 1 and Speaker 2 rather than names, unless you map them yourself.
Some speech-to-text APIs include it as an option; others need a separate model or service. Quality drops with overlapping speech, similar voices, many speakers and very short turns, and some APIs let you specify the expected number of speakers. It is measured separately from word error rate.
For calls and meetings, check whether diarization works in streaming mode if you need live captions. If your system can record each participant on a separate channel, that is often more reliable than diarizing mixed audio.
Worked example
Illustrative: in a 30-minute sales call, the diarized transcript assigns 55% of speech to Speaker 1 and 45% to Speaker 2. Your app maps Speaker 1 to the sales rep from the opening greeting and reports talk time of 16.5 versus 13.5 minutes.
Common mistakes
- Assuming speaker labels stay consistent across separate recordings or API calls.
- Not using separate channels when the phone system can record each side on its own.
- Expecting real names; you need your own step to map generic labels to people.
See it on a real card: Feature support on Veo 3.1.