Sortformer

Sortformer is a diarization model built for streaming. Where the common pipeline (segmentation plus speaker embeddings) analyzes a finished recording, Sortformer keeps a running speaker cache and emits labels every fraction of a second, which is what live speaker names in a meeting window would need. The published model has four speaker slots, so a fifth voice has to share one.

Streaming diarization is still markedly less accurate than running the same job once the recording is complete, and on a real multi-person call that gap is easy to hear. Mac Note Taker keeps the live view simple - you and the other side of the call - and does full speaker separation after you stop, where it can use the whole recording and get the speakers right.