Sam Witteveen · 2026-09-23 · notable
Sam Witteveen — 'Nemotron 3 Diarization - Who Said That?'
Sam Witteveen's 23 September 2026 video covers NVIDIA Nemotron 3 Diarization, released the same day: a 100M-parameter open model that labels who is speaking, for up to eight speakers, live or on recordings.
Sam Witteveen looks at NVIDIA's new open model for answering 'who said that?' in a conversation.
What is it?
'Nemotron 3 Diarization - Who Said That?' went up on Sam Witteveen's channel on 23 September 2026, the day NVIDIA released the model. Nemotron 3 Diarization is a 100M-parameter open-weight model that splits audio by speaker, for up to eight people, in streaming or offline mode.
How does it work?
NVIDIA's model follows the Sortformer approach: it numbers speakers in the order they first talk, so labels stay stable from one audio chunk to the next. A 31-layer Transformer encoder reads 16 kHz audio, and latency can be set from 30.4 seconds down to 0.32 seconds. NVIDIA reports a 14.72% error rate, #1 on VoiceArena's Diarization-Bench.
Why does it matter?
Speaker labels are the missing piece in most meeting notes and call transcripts. The model adds them next to an existing speech-to-text system rather than replacing it, and NVIDIA says it cuts diarization error by about 40% versus its earlier four-speaker Streaming Sortformer.
Who is it for?
engineers building transcription and meeting tools