What is speaker diarization? One microphone, many voices

Upload a recording made on one microphone and the transcript comes back with the lines grouped into separate speaker tracks. Upload one that kept each side of a call on its own channel and you get separate speakers too. Those two results look identical on the page and they are not the same kind of statement, and this guide is about the difference — because a transcript that hides it is more dangerous than one that separates nothing at all.

What speaker diarization is

Speaker diarization is the standard name for software that answers "who spoke when" from the audio alone. It segments a recording, groups segments by how similar the voices sound, and assigns each group an anonymous index. It is a real and useful technique — meeting-notes tools lean on it heavily — and it is an estimate end to end: the boundaries are estimated, the number of speakers is estimated, and the assignment of a segment to a voice is estimated.

The conditions that degrade those estimates are well documented in the field's own literature: people talking over each other, similar-sounding voices, speakers at different distances from the microphone, and noisy rooms. Those are not corner cases in recorded evidence. They are the ordinary texture of it.

Separation you can rely on comes from the recording

There is a second way a transcript can tell speakers apart, and it involves no inference at all. Many recording systems — phone platforms in particular — write each side of a conversation to its own channel in the file. When that is true, "this passage came from that side" is a fact about the recording itself. Nothing is being guessed; the file arrived already separated.

That is the distinction that decides everything here. Channel separation is a property of the file. Diarization is an opinion about the file. The two can look identical on a printed page, which is precisely the problem.

What Pincite Audio does

Where the recording keeps speakers on separate channels, the transcript uses that and does not diarize. The labels describe the channels — a role, one side of a call — and they stay descriptions until a person attests to more. The exact label shapes are documented on the format page, which is checked against the code rather than maintained by hand.

Where it does not — a single microphone in a room, a hearing, an interview — the recogniser groups the lines by voice and the transcript shows those groups as separate tracks. Three things are said about them, on the page rather than in a footnote: the tracks are machine-derived, how many people spoke is not established, and no track carries a name until a person puts one there. Naming a track from its label records the name as your identification.

There is one shape the product declines to attempt. Long recordings are transcribed in parts, and the grouping is only ever valid inside one part — a track in one part and a track in another are separate estimates about separate stretches of audio, with nothing connecting them. Rather than join them and imply a continuity nobody established, a recording long enough to be split keeps the single-speaker attribution and says so.

Why the estimate is printed, and why it is marked

In a meeting summary, a diarization error costs a mislabeled paragraph. In evidence work, an attribution is often the entire point of the exercise — who said the sentence matters at least as much as what was said. That cuts both ways, and it took a real transcript to see the second edge: a hearing with a judge and two counsel, printed as one speaker from beginning to end, is not a cautious transcript. It is a confidently wrong one. Nothing on that page was estimated, and the structure it showed was still false.

So the estimate is printed and the typography does the work the caution used to do. A machine-derived track never carries a proper name, the transcript states which mechanism separated its speakers, and the count of tracks is never presented as a count of people. This is the same rule as the one that keeps proper names off machine guesses, applied to structure instead of to identity: a statement the record cannot support does not get to look like one it can — and refusing to make the statement at all is not the same as making it honestly.

What to do with this

If you have a choice at capture time, choose the path that records speakers separately — for calls, a system that writes each side to its own channel turns speaker separation from an estimate into a property of your evidence, and it is the only version of this that needs no checking. If the recording is already made and it is one microphone, read the speaker tracks as a first pass: check the turns that matter against the audio, name the tracks you can attest to, and let each label's recorded provenance say who made that call.