NVIDIA releases Nemotron 3 Diarization: who spoke when, up to eight voices, in a single 100M model
In one sentence On September 23 NVIDIA published Nemotron 3 Diarization, a 100-million-parameter open-weight model that separates up to eight speakers in live or recorded audio, with a single checkpoint for streaming and offline. OpenMDW 1.1 license; the gated Hugging Face preview is, however, limited to internal evaluation.
Diarization is the task of working out "who spoke when" in a recording: it powers meeting minutes, call centers, multi-speaker subtitles. Until now it was usually done with a chain of models: one cuts the audio into segments, another computes a voice fingerprint, a third groups the fingerprints.
Nemotron 3 Diarization does it all in one pass. It is small, about 100 million parameters, and tracks up to eight people at once, both live and on recorded files. Speakers are numbered in the order they first speak.
The model does not transcribe words or recognize people by name: it has to be paired with a speech recognition system. Usage terms need checking: the model card says it is commercially usable, but one outlet reports that the Hugging Face preview is restricted to internal evaluation.
Companies
NVIDIA
Tools
Nemotron 3 Diarization, Streaming Sortformer
Tags
Sources