Speaker diarization
Who spoke when
Splitting a recording into turns labeled by speaker, answering "who spoke when" without necessarily knowing who the speakers are.
In DocSlurpGlossary entry
On this page
Speaker diarization splits a recording into turns and labels each turn by speaker: who spoke when. It does not identify people. It clusters stretches of audio by voice and gives each cluster a label such as “Speaker 1” or “Speaker 2”.
A typical system finds where speech occurs, turns short windows of speech into voice embeddings (vectors that are close for the same voice), and clusters them. The output is a list of turns with start and end times, which is then merged with a speech-to-text transcript by matching times:
[03:02–03:09] Speaker 1: So where does guidance land after a quarter like that?
[03:10–03:14] Speaker 2: EBITDA guidance rises to ten point six million this year.
For search over investor-call.mp4, the labels let a reader see that the answer came from a different person than the question, and let a summarizer attribute decisions to someone instead of to “the call”.
The gotcha: labels are consistent only within one diarization pass. Speaker 2 in this recording has nothing to do with Speaker 2 in next week’s call, and if a long recording is diarized in pieces, Speaker 2 can become Speaker 1 halfway through. Diarize a recording in one pass when you can. Mapping labels to real names needs separate evidence: an enrolled voice sample, a calendar invite, or someone saying “this is Dana”. Overlapping speech and short interjections (“yes”, “right”) are where diarization errs most.
In DocSlurp
In DocSlurp
When remote media processing is enabled, one diarization request covers the whole recording so labels stay consistent, and turns become transcript lines such as [03:10–03:14] Speaker 2: ... (core/worker/src/media/diarize.ts, core/worker/src/media/canonical.ts). A diarization failure leaves the transcript unlabeled instead of failing ingest.