NVIDIA has released Nemotron 3 Diarization, an open-weight, 100M-parameter model for real-time speaker diarization that supports up to eight speakers. The model ranks #1 on VoiceArena's Diarization-Bench leaderboard with a 14.72% Diarization Error Rate (DER), outperforming the next-ranked system by approximately 24% relative reduction.
- Supports up to eight speaker channels and handles overlapping speech using arrival-time ordering for stable speaker labels.
- Achieved a 14.72% DER on VoiceArena's benchmark, compared to 19.3% for the second-place system.
- Trained with public data and licensed multispeaker conversations from David AI, which decreased compound DER by 0.77 absolute points.
- Offers configurable streaming latency with recommended input-buffer latencies of 30.4, 1.04, 0.64, and 0.32 seconds.
- Demonstrates an average 41.0% relative DER reduction across eight public benchmarks compared to NVIDIA's previous four-speaker Streaming Sortformer baseline.
The model enables applications to create speaker-attributed transcripts by combining diarization timestamps with automatic speech recognition, improving the utility of conversation analytics and voice-agent memory.