Speaker diarization vs speaker identification

Two terms get used interchangeably in the transcription world, and the confusion costs people time. Speaker diarization and speaker identification are different problems: one finds the voices, the other names them. Knowing which one you are asking for explains why a tool can split a transcript perfectly and still leave it full of "Speaker 1" and "Speaker 2".

Diarization: who spoke when

Diarization takes one audio track and answers a structural question: where does the speaker change? It clusters the speech into voices — Speaker 1, Speaker 2, Speaker 3 — and assigns every word to a cluster. The output is a transcript with turns, but the labels are anonymous by design: the model grouped acoustic patterns, it did not recognize anyone.

That anonymity is a feature at this stage. Diarization has to work on recordings of strangers, and grouping by voice is something acoustics can do: two distinct voices cluster cleanly, a voice before and after a microphone change stays together (most of the time), and short backchannels get attached to the right cluster (some of the time). It also fails in characteristic ways — over-splitting one voice into two, folding a quick interjection into the neighboring speaker, and struggling with crosstalk — which is exactly why good diarization ships with a review flow instead of a confidence percentage.

Identification: which real person is that

Identification takes the anonymous clusters and answers a different question: who is Speaker 2? There are three ways this happens in practice:

• Manual mapping. A human reads enough context to know the second voice is Maya and renames the label. For interviews, panels and podcasts this is the normal path — the mapping is usually a handful of decisions for the whole recording.
• Participant lists. Meeting tools identify speakers by who was in the call, not by their voice. It works because the roster already exists.
• Voice-prints. A biometric system compares voices against enrolled reference audio. This is real technology with real prerequisites — consent, enrollment recordings, degraded accuracy on noise — and it is the only "automatic identification" that exists for recordings of people you cannot ask.

What identification is not: it is not something the acoustics of a recording can give you for free. The recording knows there are two voices; it does not know their names.

Where the boundary sits in practice

If you transcribe a two-person interview, the honest pipeline looks like this: automatic diarization produces the turns, then you identify — rename Speaker 2 to Maya, merge the cluster that over-split, reassign the one line that got folded into the wrong voice. That second stage is human work, but it is bounded: the review queue counts the flagged cues, and the whole identification pass for a ten-minute interview is typically a few clicks.

This is the boundary we drew in our speaker labels: diarization is automatic and included with every transcription; identification stays yours. No voice-prints are taken and no biometric data is processed — the mapping from "Speaker 2" to "Maya" is a human decision, applied through rename, merge and reassign, and it persists with your draft.

The one-line version

Diarization is a clustering problem and machines do it; identification is an identity problem and, for recordings of real people, a human closes it. Tools that conflate the two either silently take biometric liberties or promise a naming they cannot deliver. A transcript pipeline that does the first automatically and makes the second a two-minute review pass gets you a quotable, correctly attributed transcript without either problem.