True diarisation is an acoustic problem: the system clusters the audio by voice characteristics and assigns turns to speakers, independent of what was said. It handles overlapping speech and multi-party calls, and it produces word-level timings.
A cheaper approximation is to infer speakers from context while transcribing, who is asking questions, who is giving account details. It reads correctly nearly all the time on a two-party call and does not need a separate acoustic model.
In QXAI
We use the context-based approach, not acoustic diarisation. On a straightforward two-party call the labels are reliable; on a heavily overlapping or multi-party call they are less so. We also do not produce word-level timestamps.