Speaker diarization — identify and separate multiple speakers in audio.

Synopsis


Description

Analyzes audio to identify different speakers and when they spoke. The output includes a transcript with speaker labels.

Arguments


Options


Examples

Basic diarization

With known speaker count

Transcript output (default behavior)

JSON output

Single-model diarization

Nemotron-3.5-Transcribe-30B-A4B-FP8 transcribes, labels speakers and times words itself, so no diarizer, ASR or aligner model is loaded. Do not pass --asr-model with it: the server returns 400. --num-speakers can be at most 8, and only adds a warning when the model counts a different number. The model needs about 32 GiB of memory; see Single-Model Diarization.

Full pipeline with custom models


Output Formats

Text

JSON

Verbose JSON (with transcription)

words is shortened here; the real output lists every word.

Available Models


See Also