Nvidia has released Nemotron 3 Diarization, an open-weight model that marks when each person speaks in live or recorded audio. The 100-million-parameter system supports as many as eight speakers, including overlapping speech, and can feed its timestamps into speech recognition to build attributed meeting, call or podcast transcripts.
The model assigns anonymous channels in the order voices first appear and keeps those assignments stable across audio chunks with a speaker cache and recent-frame memory. It does not identify real people. Applications must connect labels such as speaker_2 to participant names through separate metadata or verification, and speech recognition is still required to produce words.
In the initial VoiceArena Diarization-Bench, Nemotron ranked first among 12 systems and 17 configurations, recording a 14.72% diarization error rate across about 22 hours of English conversations. Nvidia reports lower error than its earlier four-speaker model across shared latency settings. The benchmark is still undergoing paired statistical analysis, and one two-speaker subset favored the older baseline.
Developers can choose input-buffer latencies from 30.4 seconds down to a recommended minimum of 0.32 seconds, trading context for responsiveness. Those figures exclude computation, networking and transcription. Noise, distant microphones, domain shifts, long recordings or more than eight speakers can still produce missed speech and incorrect assignments.