A voice label is not a person’s identity
Diarization groups speech by voice and timing. Labels such as speaker_0 and speaker_1 do not identify a guest by name. Listen to the original recording before replacing them with names.
Review the boundaries
Short replies, interruptions, crosstalk and similar voices are common error cases. Check the first few turns, every apparent new speaker and any passage where the wording or label looks surprising.
- Enable speaker labels before transcription.
- Compare each label with the recording.
- Edit names, text and timing in the transcript workbench.
- Export a reviewed transcript or caption file.
Separate tracks can be simpler
If each microphone already has its own channel, splitting the channels gives you a known source for each speaker. This is different from separating voices that have already been mixed together.
The workbench requests real provider labels and shows a warning when they are absent. Missing labels are not evidence that only one person spoke.