What is speaker diarisation?
Speaker diarisation is the separation of an audio stream into segments by speaker, so that a transcript records who said which part. On a two-party telephone call it answers whether a given sentence came from the agent or from the caller. That attribution is what turns a transcript into evidence, since a disclosure only counts if the agent is the one who made it.
Also called: Speaker segmentation
What goes wrong when a transcript has no speaker labels
Diarisation is often assumed to be solved on a telephone call, since there are only two parties and often two separate audio channels. The assumption holds until a carrier mixes the channels, or until a third person joins the line at the caller's end. At that point the transcript still reads plausibly while attributing the agent's disclosure to the customer, and the compliance record is quietly wrong. Attribution errors are also concentrated exactly where people speak over each other, which is where disputes originate.
- A disclosure only counts in evidence when it is attributed to the agent.
- Mixed mono recordings remove the channel separation the platform relied upon.
- Errors concentrate on overlapping speech, which is where disputes originate.
- Analytics computed per speaker are wrong wherever attribution is wrong.
How speaker diarisation works
- Channel separation is preserved from the carrier wherever the recording allows it.
- Speech regions are detected and divided into short homogeneous segments.
- Each segment is converted into a speaker embedding describing the voice.
- Embeddings are clustered to group segments belonging to the same speaker.
- Clusters are mapped onto known roles using the channel or the call flow.
- Overlapping regions are flagged rather than assigned to a single speaker.
How speaker diarisation is evaluated
Diarisation is measured on real recordings, including the ones with overlap and a third party present. Examine the following across sampled calls:
- diarisation error rate including overlapping speech
- speaker confusion rate between agent and caller
- accuracy of role assignment to each cluster
- performance on mixed mono against dual channel recordings
- detection rate for a third speaker joining mid-call
Limitations, risks and common failure modes
- Carrier recordings arrive mixed to mono, which removes the easiest separation signal.
- A supervisor joining the call creates a third cluster the role mapping does not expect.
- Overlapping speech is assigned wholly to one party, misattributing the other's words.
- Compliance reporting counts disclosures from transcript labels nobody has validated.