NVIDIA Releases Nemotron 3 Diarization: 100M-Parameter Open-Weight Model for 8 Speakers

NVIDIA has released Nemotron 3 Diarization, a 100M-parameter open-weight speaker diarization model on Hugging Face. It tracks up to 8 overlapping speakers and handles both offline recordings and real-time streaming under the commercial-friendly OpenMDW 1.1 license.

• 100M-parameter model tracks up to 8 speakers, including overlapping speech
• One checkpoint covers offline recordings and streaming with latency from 0.32 s
• Ranked #1 among 12 systems on Voice Arena Diarization-Bench at 14.72% DER
• Trained on ~10,000 hours of real conversations and 82,611 hours of simulated mixtures