NVIDIA has published Nemotron 3 Diarization, a small open-weight model that decides who spoke when, including when several people talk over each other.
The model carries about 100M parameters and tops VoiceArena’s Diarization-Bench on its first public run, posting a 14.72 percent diarization error rate across roughly 22 hours of English conversation. The next system back sits at 19.3 percent.
Speaker labelling follows arrival order, an approach NVIDIA inherited from its earlier Sortformer work. The first new voice takes channel one, the next takes channel two, and so on, which keeps labels stable as audio streams in. Support extends to eight speakers, double the previous four.
Against NVIDIA’s own four-speaker streaming baseline, the new model cuts error by an average of 41 percent relative at 1.04 seconds of buffered audio, with gains widening in meetings that carry more participants. Training mixed public and licensed speech, including real conversations from David AI that contributed multilingual material spanning 21 languages.
Accuracy is not uniform. On the two-speaker CALLHOME subset the older model is marginally better, while the full CALLHOME evaluation still improves. NVIDIA also reports large throughput jumps, measured at batch size 32 on an RTX PRO 5000.
