What is Diarization?
Marking which speaker said which segment in a multi-speaker recording.
Diarization and identification are not the same thing: diarization only answers that the speaker changed, identification answers who it changed to.
In multi-speaker conversation data, whether diarization is required changes the labour by a multiple, which is why it is one of the first questions in any quote.
Levels of diarization output
The basic output is a set of speaker turns: speaker 1 from the start to eleven seconds, speaker 2 from there to nineteen. The labels are local to the recording. Speaker 1 in one file has no connection to speaker 1 in the next, unless identification is layered on top. Buyers often assume the labels line up across files, and they do not.
The second distinction is offline against online. Offline diarization sees the whole recording before deciding who spoke when, and can revise earlier judgements as more audio arrives. Online diarization decides as the audio streams in, with less context, and it is what live captioning and meeting assistants use. The same recording is a harder problem for the online system, and its output should be judged against that.
The hard cases separate a usable result from a rough one: overlapping speech, where two people talk at once and the segment belongs to both; very short turns, the mm-hmm and yeah a speaker drops while someone else is talking; and speakers who sound alike, such as siblings or two people on the same line. Deciding how many speakers are present is part of the task.
Where it shows up in a quote
Diarization is a line of its own, separate from transcription. The usual structure is two passes: transcribe the audio, then mark who spoke when, and join the two so each transcript segment carries a speaker label. A buyer who wants a speaker-attributed transcript is buying both passes, and a quote that names only one of them will grow later.
The labour scales with the number of speakers and the amount of overlap, not with hours alone. A clean two-person interview is the easy case. A five-person meeting with crosstalk is several times the work for the same hour of audio. Whether the client needs anonymous turns, the cheap version, or names attached to each speaker, which needs enrolment samples or a reference list, changes the price again.
On the specification side: the minimum turn length worth marking, whether overlapping speech is labelled as overlap or assigned to one speaker, and the output format. The common interchange format is RTTM, a plain text file listing a start time, a duration and a speaker per segment, which most scoring tools read. On delivery, check one file by ear against its labels.
Related terms
-
Speaker Identification
Identifying who is speaking from the voice itself. Also called voice biometrics.
-
Conversational Speech
Speech from natural conversation between two or more people, on open or semi-structured topics, used for conversational AI, voice assistants, and small talk models.
Keep reading
-
Full glossary
Every term we explain, in one list.
-
How to buy training data
Where these terms actually show up, and which ones change a quote.
-
Help center
How a project runs, from specification to delivery.
Submit a sourcing request
Tell us the language, the hours, and what the data needs to look like. You will get a real number and a real timeline, not a range. If we cannot source it well, we will tell you that instead.
- Pilot batch before the full run, so problems surface early.
- Consent documentation delivered with the data.
- No medical or clinical data. No recorded telephone calls.
- You send the spec Language, hours, recording conditions, and the deadline.
- We reply within two business days A real number and a real timeline, not a range.
- You approve a pilot batch A small run first, so problems surface before the full order.
- Full run, then delivery with documents Audio, transcripts, and the consent record for every speaker.