Speaker diarization

Who spoke when is segmented across the recording, giving a turn-by-turn structure without knowing anyone's identity. It is what turns a flat transcript into something usable for meetings, interviews and calls with several participants. The number of speakers does not need to be known in advance, and overlapping speech is marked rather than assigned to one voice. Segments are returned with timings, so any turn can be located in the recording.

The service runs on a private, dedicated cluster sized for your volume — a single GPU for a pilot, up to a hundred for a production estate. It is deployed where your data already lives: on premise, in your own AWS, GCP or OCI account, or in the super-gpu cloud.