NVIDIA has released Nemotron 3 Diarization, an open-weight speaker diarization model available on Hugging Face under the OpenMDW License 1.1. The 100M-parameter model tracks up to 8 overlapping speakers in real-time or offline modes using a single checkpoint.

  • Doubles the speaker limit from NVIDIA's previous Streaming Sortformer, which supported 4 speakers.
  • Achieved first place on Voice Arena’s initial Diarization-Bench with 14.72% DER, a 24% relative reduction over the next-ranked system.
  • Offers four latency operating points ranging from 30.4 seconds for offline processing to 0.32 seconds for ultra-low latency streaming.
  • Trained on approximately 10,000 hours of real conversations and 82,611 hours of simulated multi-talker mixtures.
  • Runs on NVIDIA GPUs via NeMo and supports formats including .wav, .flac, .opus, and .mp3.

The model enables meeting tools, call analytics, and voice-agent memory systems to attribute speech to specific speakers in complex audio environments.