Speaker diarization

Diarization answers the question of who spoke when. A transcription service that supports it groups segments by voice and labels them, typically Speaker A, Speaker B, and so on. Some services can map labels to real names if you provide short reference clips; otherwise naming is up to you or the user.

It's what makes meeting notes, sales-call summaries, interviews, and support-call reviews useful. Without speaker labels, a summary can't say who committed to what, and a language model asked to infer speakers from unlabeled text will guess, sometimes confidently and wrongly. With them, you can attribute action items, measure talk time, and show the transcript as a conversation.

Support varies by model and vendor, and it's a common reason teams pick a speech specialist over a general provider. Check the docs for the exact model you plan to use rather than assuming a newer model keeps every feature of the one it replaces. Diarization degrades with overlapping speech, similar voices, and many speakers, and labels can swap partway through a recording. Make labels editable in the UI, and when you summarize, pass the labels through so the model attributes statements correctly.