Nvidia releases Nemotron 3 Diarization, a speaker-diarization model that tops the benchmark

Nvidia has released Nemotron 3 Diarization, a speaker-diarization model that identifies who spoke and when even when voices overlap, supports up to eight speakers, and weighs in at 100 million parameters. The model is available now on Hugging Face.

On voicearena_ai's Diarization-Bench, the model ranked first out of 12 systems with a 14.72% error rate, about 24% lower than the second-place entry. The gap is striking given the model's compact size.

Overlapping speech remains a hard problem in speech processing; most systems break down when two people talk at once. Nemotron 3 maintains consistency even in scenarios with four simultaneous speakers, as demonstrated in the published benchmark comparison.

Availability on Hugging Face lets developers plug the model into transcription pipelines, conversation-analysis tools, and accessibility applications without building their own training infrastructure. At 100 million parameters it is light enough to run on a single GPU, lowering the barrier for small teams. The results come from an initial version of the benchmark, and Nvidia has not published additional metrics such as latency or memory consumption in practical deployment. Independent runs are worth watching before committing to production.