Diarization is the audio-processing task of segmenting a recording into speaker turns: 'speaker A said this, then speaker B, then speaker A.' On a one-on-one call it is trivial; with multiple voices on the line it becomes necessary infrastructure.
The use cases are more common than vendors admit. A home-services call where the homeowner and the spouse are both on the line. A clinic call where the parent and the child are talking. An auto-service call where the customer hands the phone to a partner mid-conversation. Without diarization, the agent attributes everything to a single caller and gets the context wrong.
Production-grade diarization is fast (sub-200 ms decision latency) and stable (it does not flip between speakers turn by turn). It also has to interact correctly with voice authentication, the verified speaker is one of the two voices on the line, and the unverified spouse asking for the appointment change should not be granted the verified caller's authority.
Vorel ships diarization on every call, with speaker-tagged transcripts and a verification policy that does not silently extend authority across speakers.