NVIDIA's 100M Model Tracks 8 Speakers in Real Time

NVIDIA put a new speaker diarization model on Hugging Face this week, and the numbers are small on purpose. Nemotron 3 Diarization runs on 100 million parameters, tracks up to eight speakers at once, and handles overlapping voices in the same pass, whether the audio comes in as a finished recording or a live stream. That last part matters: one checkpoint covers both offline and real-time use, so teams don't need separate models for a podcast archive and a live call center feed.

The problem it's solving is basic but stubborn. Speech recognition can transcribe every word in a meeting, but it has no idea who said which line. That gap breaks anything downstream that depends on attribution, a summarizer that needs to know who agreed to what, a call analytics tool tracking which rep said which phrase, a voice agent trying to remember who it's talking to. Diarization fills that gap by marking exactly when each speaker is active, and those timestamps get layered onto the transcript afterward.

NVIDIA released the weights under the OpenMDW License 1.1, which allows commercial use, and built the release to actually run in production rather than sit as a research demo.