The AI Post
Agents & CodingOpen ModelsEnterpriseFundraisingGenerative MediaGovernanceInferenceInfrastructureLegal & SafetySector Impact
← Front Page Model Releases · Nvidia · Hugging Face · VoiceArena

Nvidia releases 100M open‑weight diarization model, says it tops VoiceArena

Nemotron 3 Diarization labels who spoke when for up to 8 speakers, and Nvidia says it cuts diarization error to 14.72% against 19.3% for the next system.

Nvidia released Nemotron 3 Diarization on Wednesday, a 100-million-parameter open-weight model that labels which speaker is talking in a recording or a live stream, for up to 8 speakers. The company published it in a post on Hugging Face.

Nvidia says the model ranks first on VoiceArena's Diarization-Bench, with a 14.72% diarization error rate against 19.3% for the next system. That figure is the company's own, taken from its post, and nobody else has yet published a test of the release.

The model takes single-channel audio at 16 kHz and runs offline or streaming, with input-buffer latency configurable from 30.4 seconds down to 0.32 seconds, according to the post. It carries the OpenMDW License Agreement v1.1 rather than one of the usual open-source licences.

Nvidia puts the gain at an average 40% relative reduction in diarization error at 1.04 seconds of latency, measured against its own earlier Streaming Sortformer baseline. On DIHARD III it reports 12.73% error at the 30.4-second setting. Both comparisons are the company's, on datasets it selected.

The weights sit on Hugging Face and the code ships in Nvidia's NeMo Speech framework on GitHub, with a demo Space linked from the post. The model page returned an authorisation error to an anonymous request, which may mean the licence gates access to the files.

Sources 1 source

  1. Source Nvidia