Speaker diarization splits an audio recording into segments by who is speaking. It answers the question 'who spoke when'.
It spots where the speaker changes, then groups the voice segments.
A transcript can then tie each line to the right person, even when the speakers are unknown in advance.
How diarization works
A diarization pipeline runs in stages. Voice activity detection first strips out silence and non-speech.
The rest of the audio is cut into short segments. Each segment becomes a speaker embedding.
That is a number-based fingerprint of the voice. A clustering step then groups segments with similar fingerprints.
It gives them plain labels such as Speaker 1 and Speaker 2.
It also guesses where one speaker stops and the next starts.
Newer systems do this end to end with neural networks. They handle speech that overlaps better than older clustering.
Diarization can run on a finished recording, or step by step on a live stream.
A finished recording gives the model the whole conversation to work from, so it has the most context.
Diarization vs speaker identification
Diarization splits voices apart. It does not know whose they are. The output is 'Speaker 1 said X'.
Speaker identification goes one step more. It maps each voice to a named person. It can match against enrolled voice profiles.
In video meetings it can use signals from the platform. Examples are active-speaker events and participant data.
Meeting products usually mix the two. A transcript then reads 'Priya: X' and not 'Speaker 1: X'.
Why diarization matters
A transcript with no names is hard to act on. Attribution is what makes talk-time stats work.
The same goes for coaching notes, action-item owners and solid summaries. It matters most on calls with several speakers.
Think sales calls, interviews, panels and research sessions. There, knowing who asked and who answered changes what the text means.
Why does diarization mislabel speakers?
Diarization from one mixed audio track still makes mistakes. Short back-channel phrases can throw it off. So can similar voices and crosstalk.
Segments then merge or swap. Video meetings offer a shortcut.
When each person's audio arrives on its own stream, attribution is all but exact. Each track already belongs to a known person.
You then need it only inside a track, when several people share one microphone in a room.
Diarization in practice
MeetStream sorts out who is speaking at the capture layer. On Zoom each participant gets an isolated track.
On Google Meet and Teams the audio is speaker-attributed by platform identity. Transcripts also carry per-word speaker labels from the engine.
Diarization then acts as a second opinion, not the only source. Details are in the MeetStream docs.
See the meeting bot API page to go deeper.
Frequently asked questions
What is the difference between diarization and transcription?
Transcription turns speech into text. Diarization works out who spoke each part. They are two steps that usually run together.
The result is a transcript. Every line carries a speaker label and timestamps.
How accurate is speaker diarization?
Take clean, single-channel audio with a few speakers. Good systems then reach diarization error rates in the single digits to low teens.
Accuracy falls with speech that overlaps, many speakers and noisy recordings. It climbs sharply when audio is captured per participant.
Can diarization run in real time?
Yes, but with tradeoffs. Streaming diarization must label speakers without seeing what comes next. Labels can change as more speech arrives.
Systems that get one audio stream per participant avoid this.
Put the vocabulary to work
One API to join, record, stream, and transcribe meetings on Zoom, Google Meet, and Microsoft Teams.