Speaker diarization

What it is

Diarization answers "who spoke when?" on a single audio stream that contains more than one voice. It clusters segments by voice characteristics and assigns a label to each. It is a distinct task from speech recognition, with its own error rate (diarization error rate) and its own failure modes: two similar voices merged into one, one speaker split into two, overlapping speech mislabelled.

On a live agent call, the caller and the agent usually arrive on separate audio channels, so the platform already knows who is speaking and diarization is not needed. It becomes relevant when a third party joins (a family member, an interpreter, a human agent after transfer), when you analyse mono recordings from your existing call-recording system, or when a warm transfer produces a three-way segment.

Why it matters when buying

Buyers meet diarization in two places. In analytics and quality assurance, where the vendor promises "agent talk time" and "customer sentiment" from recordings, the numbers depend on diarization accuracy. In compliance, where an all-party recording notice or a consent statement must be attributed to the right person, a diarization error is an evidence problem. It also affects a bake-off: if you evaluate speech accuracy on mono call recordings, diarization errors leak into word error rate.

What to ask

Ask whether the agent's own calls are recorded per channel or mixed. Ask how analytics on historical mono recordings handle overlapping speech and similar voices. Ask for diarization accuracy on your own recordings if analytics are part of the purchase. For a two-party agent, ask the vendor to confirm diarization is not in the live path adding latency.