What is speaker diarization?

Ask a transcription tool for "the speakers" and you are really asking for two things: find the voices, and tell me who they are. Speaker diarization is the first part — and knowing exactly what it does (and does not do) explains both its value and its limits.

The definition

Speaker diarization answers one question about a recording: who spoke when. It takes a single audio track and splits it into its voices, attributing every spoken word to an anonymous label — Speaker 1, Speaker 2, and so on — with the turn changes marked on the timeline. A two-person interview comes back as a dialogue instead of a wall of text; the words were already right, the attribution is what diarization adds.

The labels are anonymous by design. Diarization is pattern grouping, not recognition — it never learns anyone's name. Mapping "Speaker 2" to a real person is a different problem, speaker identification, and it happens after the diarization, not inside it.

What the output looks like

A diarized transcript gives every cue three things it did not have before: an attribution, a color, and a place in a turn structure. In SubtitleGenerator's editor that means a legend listing every voice found, a color dot on each subtitle row, a marker on the timeline at every turn, and — for the lines the model is not sure about — a flag, with a counter that says how many cues need a speaker check.

When it works well, and when it does not

Diarization is strongest where voices are distinct and turns are clean: a two-person interview with good microphones, a panel clip where people take turns, a structured conversation. It is weakest where the acoustic patterns blur — background noise, crosstalk, very short interjections, and long recordings where a voice drifts (microphone changes, distance, fatigue).

Those failures have names: over-splitting (one voice becomes two labels), under-splitting (two voices share one label), and single-line misattribution (a backchannel lands on the wrong speaker). No honest tool claims these never happen. The useful question is what happens next: a transcript with flagged, counted, reviewable cues turns the failures into a bounded cleanup instead of an unbounded rewrite.

What it is for

Any recording with more than one voice benefits — interviews, video podcasts, webinars, lectures, meetings. If your footage has one voice, diarization has nothing to add, which is why editors show no speaker surfaces at all for single-voice results.

In SubtitleGenerator, diarization runs automatically as part of transcription on every plan, and the correction pass — rename, merge, reassign, confirm — is built into the same review flow as the flagged words. The capability page shows the full flow; the identification guide covers how the anonymous labels become real names.