Published: 24 September 2026
NVIDIA published Nemotron 3 Diarization on 23 September 2026, a 100 million parameter open model that works out who spoke when in a recording. It scores a 14.72 percent diarization error rate on Voice Arena’s Diarization-Bench, first out of the twelve systems tested. NVIDIA released a second model the same day, NV-Reason-CT, which reads 3D CT scans and writes structured radiology reports.
What does Nemotron 3 Diarization do?
Speaker diarization is the step between a recording and a usable transcript. Speech recognition gives you the words. Diarization tells you which of the people in the room said each of them. Without it, a six person meeting transcribes as one long block of text.
This model handles up to eight speakers, including people talking over each other, and runs either after the fact or live. NVIDIA ships four latency presets, from a 30.4 second buffer for offline batch work down to a 0.32 second buffer for real-time captioning. The accuracy gap between those two extremes is under one point, which is the interesting part.
What you need to run it:
- Licence: OpenMDW 1.1, which permits commercial use
- Download: nvidia/Nemotron-3-Diarization on Hugging Face
- Hardware: one NVIDIA GPU, Ampere generation or newer, on Linux
- Software: the NVIDIA NeMo toolkit
- Input: 16 kHz mono audio as .wav, .flac, .opus or .mp3
- Languages: English, Mandarin, Hindi, Kannada, Telugu and Bengali
Nemotron 3 Diarization benchmarks and latency
Take this from the table: the model is well ahead of the field on the headline test, and the speaker count is what moves the number, not the latency setting.
| What is measured | Nemotron 3 Diarization | Comparison |
|---|---|---|
| Diarization-Bench error rate (lower is better) | 14.72% | 19.3% for the next best of 12 systems |
| DIHARD III error rate, 1 to 4 speakers | 9.13% | rises to 27.58% with 5 to 9 speakers |
| Live mode, 0.32 second buffer | 13.55% | 12.73% offline with a 30.4 second buffer |
| Speakers handled | up to 8 | Streaming Sortformer, its predecessor, handles 4 |
| Speed | 15,113x faster than real time at batch size 32 | one GPU, Ampere or newer |
| Licence | OpenMDW 1.1, commercial use allowed | weights published on Hugging Face |
The Diarization-Bench figure is Voice Arena’s, measured on 139 English conversations totalling about 22 hours. The DIHARD III numbers are NVIDIA’s own, published in the model card. NVIDIA also reports an average 41 percent relative error reduction against Streaming Sortformer across eight evaluation conditions. Training used roughly 10,000 hours of real conversation plus 82,611 hours of simulated multi-speaker audio across 21 languages.
What is NV-Reason-CT?
The second release is a vision-language model for radiology. NV-Reason-CT takes a full 3D chest or abdominal CT scan and produces a structured diagnostic report, showing its reasoning the way a radiologist writes one. It supports follow-up questions rather than a single pass.
- Built on Qwen3.5-4B as the language backbone
- CT-RATE benchmark: macro-F1 of 0.614 and macro-AUROC of 0.871
- Beats VoxelFM, Pillar-0, CT-CLIP, Merlin and ClinFusion-8B on NVIDIA’s figures
- Checkpoints on Hugging Face, with inference scripts and training configs on GitHub
- The launch post does not name a licence, and the model has no regulatory clearance
What this means
Nemotron 3 Diarization is worth testing now, and it is the more useful of the two. A four or five person team building meeting transcription, call centre analytics or podcast tooling has been stitching together a recognition model and a separate diarization step that quietly falls apart once a fourth person joins. A 100 million parameter model that fits on a single older GPU under a licence that allows commercial use removes both the accuracy problem and the hosting bill. Run it on your own recordings first, because the published gains come from English conversations and NVIDIA has not broken the error rate down by language despite claiming six.
The number to check before you commit is the 27.58 percent error rate on recordings with five to nine speakers. That is three times the figure for smaller groups, and it sits under a headline that says eight speakers. If your recordings are two-person sales calls you will see the good number. If they are workshops, you will not. NV-Reason-CT is a different proposition: interesting research, published without a licence and without clearance from any regulator, so it belongs in a lab and nowhere near a patient.
For more information, visit the Nemotron 3 Diarization model card on Hugging Face.
