Nvidia has released Nemotron 3 Diarization, a roughly 100-million-parameter model that identifies who is speaking at each moment in recorded or live audio. Its weights are freely available, giving developers a compact component for meeting transcription, call analysis and voice-agent pipelines.
Speaker diarization means dividing audio by speaker rather than converting speech into words. This model distinguishes as many as eight people and can detect overlapping speech. Paired with a speech recognizer such as Nvidia's Parakeet, it can produce a transcript with anonymous labels such as “speaker_2”; identifying the actual person requires a separate step.
Developers can choose audio buffers from 30.4 seconds down to 0.32 seconds to trade delay against accuracy. Shorter buffers generally make more errors. Additional speakers, strong background noise and reverberant rooms also degrade the result.
On VoiceArena's Diarization-Bench, Nemotron 3 recorded a 14.72% error rate, compared with 19.3% for the next system. The benchmark counts overlapping speech and small timing mistakes at speaker changes. At a 1.04-second buffer, Nvidia reports an average error reduction of 41% across eight scenarios versus its earlier Streaming Sortformer model.