NVIDIA Nemotron 3 Diarization: real-time diarization for up to eight speakers
Open-weight, 100M-parameter model for speaker diarization supporting up to eight speakers, ranked #1 on VoiceArena Diarization-Bench with a 14.72% DER.
What changed
Hugging Face published NVIDIA Nemotron 3 Diarization, an open-weight 100M-parameter diarization model that supports up to eight speakers and streaming and ranked #1 on VoiceArena with 14.72% DER.
Why it matters
Developers can add overlap-aware speaker timestamps to ASR pipelines to produce speaker-attributed transcripts and improve meeting analytics.
Who should care
DevelopersEnterprisesResearchers
Worth trying
Yes — Open-weight, 100M-parameter model with streaming and up to eight speakers.