NVIDIA releases Nemotron 3 Diarization model, ranked #1 with 14.72% error rate

NVIDIA shipped a small but state-of-the-art speech model: Nemotron 3 Diarization, which took the #1 spot on Diarization-Bench with a 14.72% error rate, roughly 24% lower than the runner-up. Speaker diarization — determining 'who spoke when' — is a persistent bottleneck for meeting transcription, call analytics, and multi-speaker voice agents, and overlapping speech is its hardest case.

At just 100M parameters, the model is deliberately lightweight, handling up to eight speakers and tracking overlapping voices, which makes it deployable on modest hardware. NVIDIA released it on Hugging Face with a live demo on Hugging Face Spaces, and paired the announcement with hardware tie-ins to Pollen Robotics' Reachy Mini and DGX Spark.

The release fits NVIDIA's broader Nemotron open-model push — the Nemotron 3 Super 120B is also available on Amazon Bedrock for coding-agent workflows — as NVIDIA builds a software and model ecosystem atop its silicon dominance. Community context is mixed: SemiAnalysis and open-source developers have flagged accelerator lock-in worries after NVIDIA's large Hugging Face deal, eyeing alternatives like ModelScope. Still, a small, top-ranked, openly available diarization model is a concrete, useful drop. What to watch: real-world robustness on noisy, accented, and far-field audio versus the benchmark.