I have been in a lot of stressful boardrooms. The pattern is always the same: someone junior sits down with a laptop, does not contribute to the conversation, and spends the hour writing down what everyone else said, because the action items land on whoever took the notes. That person's time is the tax on the meeting.
NVIDIA released Nemotron 3 Diarization on 23 Sep 2026, and it is aimed almost exactly at that person. Diarization is the who-spoke-when layer: given audio, it outputs segments like speaker_1, 0.51s, 12.62s. It does not transcribe. It attributes. Transcription models have been good for years, and the attribution was the piece that was missing from meeting capture.
The model card is the primary source I checked, and NVIDIA's HuggingFace blog sits alongside it with a headline the card doesn't carry: Nemotron 3 Diarization ranked #1 on VoiceArena's Diarization-Bench leaderboard, 14.72% diarization error rate (DER) across 139 English conversations, against 19.3% for the next-ranked system. NVIDIA's own evaluation reports an average 41% relative DER reduction at 1.04-second latency across eight benchmark conditions, ranging from 9% on CALLHOME to 65% on NOTSOFAR1.
The model itself is a 100M-parameter streaming Sortformer, open weights under the OpenMDW license, ready for commercial use. It handles up to eight speakers, overlapping speech included, and runs streaming at input-buffer latencies from 30.4 seconds offline down to 0.32 seconds ultra-low. On a single RTX PRO 5000 Blackwell it processes audio at 12.5x realtime at that ultra-low setting, and 199x realtime at batch 32.
The accuracy numbers deserve the same honesty as the speed numbers. On DIHARD III, the offline diarization error rate is 12.73% against 19.09% for the previous streaming Sortformer, which capped at four speakers. The gap holds at ultra-low latency: 13.55% versus 19.85%. The jump is largest in the five-to-nine-speaker band: 40.21% down to 27.58%. One caveat the blog flags itself: on the two-speaker CALLHOME subset, the old baseline still edges ahead, 5.68% versus 5.98%. The gains concentrate where meetings get crowded. Eight speakers is a hard ceiling, though. The model resolves speaker permutation by ordering its eight channels by first arrival, which also means a ten-person meeting gets a wrong answer by construction.
Editorial
This is the kind of AI release I have wanted to see for a long time. A small, well-scoped component that removes a specific, recurring cost from ordinary working days.
The economics are the point. Minutes-taking used to mean either a human sacrificing their own contribution, or a per-seat SaaS subscription with a monthly fee per attendee. Open weights plus a 100M model that runs on a single workstation card means a team can stand up diarization once, pair it with any transcription model they already use, and not pay per meeting or per seat. The 0.32-second latency means it can run live in the room, so the minutes draft itself as the meeting happens, and the junior person on the laptop can finally be a participant instead of a stenographer.
The honest limits: it labels speakers generically, speaker_1, not names, so identity resolution is still on you. And 13.55% error at ultra-low latency means a meeting transcript still needs a human skim before it becomes minutes anyone signs off on. The model does not write the minutes. It ends the note-taking tax, which is where most of the pain was.
Working-level people deserve this technology. The meeting never changes; the cost of being in it can.