Nvidia released Nemotron 3 Diarization, a free 100-million-parameter AI model that identifies which speaker is talking in a conversation in real time. The model can distinguish up to eight speakers, handle overlapping speech, and process both live audio and recorded files. When paired with a speech recognition system like Parakeet, it produces transcripts featuring anonymous speaker labels.
The audio buffer can be adjusted across four levels ranging from 30.4 down to 0.32 seconds, though shorter buffers reduce overall accuracy. Heavy background noise, reverb, and larger numbers of participants also increase error rates. On VoiceArena's strict Diarization-Bench, the model achieved first place with a 14.72 percent error rate, outperforming the next best system's 19.3 percent.
Compared to its predecessor, Streaming Sortformer, the new system reduces error rates by an average of 41 percent across eight test scenarios using a 1.04-second buffer. The model weights are freely available to the public. This release matters because it provides developers with a high-performing, open tool for real-time speaker identification and multi-speaker transcription tasks.