NVIDIA Nemotron 3 Diarization Achieves 14.72% Diarization Error Rate

NVIDIA's Nemotron 3 Diarization model has achieved a 14.72% Diarization Error Rate (DER), ranking first among 12 systems evaluated by VoiceArena. This model supports a maximum of eight speakers and has 100 million parameters. In comparison, the next-ranked system in VoiceArena recorded a DER of 19.3%, highlighting a significant performance advantage. The evaluation included 139 English-language conversations totaling 22 hours, demonstrating the model's robustness in real-world scenarios.

The model's performance is enhanced by incorporating licensed speech data from David AI, which contributed to a 0.77 absolute point reduction in compound DER. For real-time applications, the recommended input-buffer latency varies from 0.32 seconds for ultra-low latency to 30.4 seconds for offline processing. Notably, the Nemotron 3 model shows a 41.0% average relative DER reduction compared to NVIDIA's previous Streaming Sortformer model, which had a 60% higher error rate. This follows ongoing evaluations of the first version by VoiceArena, indicating a continuous improvement in diarization technology.