NVIDIA launches Nemotron 3 Diarization model tracking up to 8 overlapping speakers

andimarafioti · x · 2026-09-24

NVIDIA released Nemotron 3 Diarization, a 100M-parameter model that tracks who spoke when even with overlapping voices, handling up to eight speakers. It's available on Hugging Face with day-0 Transformers support, letting developers plug it straight into transcription pipelines.

Related event: NVIDIA open-sources Nemotron 3 diarization model, tops VoiceArena(11 posts)→

Original post →

More from Models

Models channel →