Nvidia's 100M-Parameter Nemotron 3 Diarization Tops VoiceArena at 14.72% Error
A.I. / news
Nvidia's 100M-Parameter Nemotron 3 Diarization Tops VoiceArena at 14.72% Error
The open model separates up to eight speakers in streaming audio, and its model card admits that Nvidia swapped the reference labels on three of its test sets.
Nvidia released Nemotron 3 Diarization on Sept. 23, a 100-million-parameter model that labels who spoke when in audio with up to eight speakers, and it now ranks first on the third-party VoiceArena Diarization-Bench with a 14.72 percent error rate. The next system, an open model from BUT FIT called DiariZen WavLM Large, scored 19.34 percent.
Nvidia published the weights on Hugging Face under the OpenMDW-1.1 licence and described the model in a blog post by seven authors including Francesco Ciannella, Ivan Medennikov and Taejin Park. The repository was not gated, and showed 48,784 downloads and 629 likes on Oct. 3.
What the model does and what it needs
Diarization is the step that decides which stretch of audio belongs to which speaker, before or alongside transcription. The model card lists a 31-layer Transformer encoder, 16 kHz mono input and eight output channels ordered by each speaker's first arrival. One checkpoint runs offline with a 30.4-second buffer or streaming at 1.04, 0.64 or 0.32 seconds. It needs a Linux machine with an Ampere, Hopper or Blackwell GPU.
The model gives anonymous labels such as speaker1. It does not identify anyone. Nvidia lists degraded results on long recordings and on audio with heavy noise, echo or distant microphones.
The VoiceArena table
Voice Arena, which is not Nvidia, scored 12 systems on 139 English far-field sessions totalling 21.97 hours, with 2 to 8 speakers per session and no speaker count given to any system. The board is marked early, and Voice Arena says rankings between close systems may change as sessions are added.
- Nemotron 3 Diarization (Nvidia)14.72 %
- DiariZen WavLM Large (BUT FIT)19.34 %
- Precision-3 (pyannoteAI)20.56 %
- Streaming Sortformer 4spk v2 (Nvidia)24.61 %
- Scribe v2 (ElevenLabs)40.71 %
- Nova-3 (Deepgram)67.13 %
Source: Voice Arena Diarization-Bench, accessed 2026-10-03
Nvidia's older Streaming Sortformer scored 24.61 percent, so the new model cuts error by 9.89 points against its predecessor on a test Nvidia did not run.
Nvidia's own results and the label caveat
On DIHARD III, a multilingual benchmark with 259 recordings, Nvidia reports its model against its previous Streaming Sortformer. These are vendor-supplied figures.
| Latency | Sortformer 4spk v2.1 | Nemotron 3 |
|---|---|---|
| 30.4 s | 19.09 | 12.73 |
| 1.04 s | 19.60 | 13.18 |
| 0.32 s | 19.85 | 13.55 |
The card's tables bold Nemotron in every row, including ones it loses. On CALLHOME-Part2 with two speakers at 30.4 seconds, the older model scored 5.68 and Nemotron 5.98. At 1.04 seconds the older model led again, 6.83 to 6.98.
The card also carries a warning that reads like a reply to a critic. Nvidia scored AMI, AliMeeting and NOTSOFAR1 against forced-alignment reference labels, because the original annotations "may label substantial within-segment silence as speech." It adds that results from other labels are "not directly comparable." AMI and AliMeeting also appear among the training sources, as simulated mixtures of 6,743 and 6,745 hours. The card does not say in the passages The Terminal read which splits of them were used.
Training data and licence
Training used about 10,000 hours of real conversations and 82,611 hours of simulated mixtures, per the card. Licensed recordings from David AI supply 36,458 hours of English mixtures and 19,216 hours across 21 languages. Nvidia says the David AI data cut the combined error rate by 0.77 points. The run used eight nodes of eight A100 80 GB GPUs.
The OpenMDW-1.1 licence is a permissive terms set for models, released by the Linux Foundation on May 28, and Nvidia said then that it plans to adopt it across Cosmos, Isaac GR00T, Ising and Nemotron. The card says the model is ready for commercial or non-commercial use. Nvidia's open-source OpenShell agent sandbox and local-inference work like the ds4 report are covered separately. Voice Arena's page carries no date for its next update, and its ranking stands as early results.
Sources
More in A.I.
- 01WorldTUI Is a 360M-Parameter Model That Fakes a Linux Terminal, and Its Weights Are Not Out YetSimon Edwardsson of V7 trained a trimmed Qwen2.5-Coder on 2 billion tokens of recorded shell sessions. It scores 80 of 106 on his own test and still forgets files it just wrote.
- 02Omnara, a 3,200-Star Apache-2.0 Rival to Claude Managed Agents, Ships With Defaults Its README Calls InsecureThe omnara-ai project runs agents on any model through a Postgres-backed event log and a daemon that dials out. Its README does not name the team behind it.
- 03Meta Paused Muse Chat Sharing After 6,800 Share Hostnames Turned Up in DNS RecordsClaroc found that Muse put each shared conversation's title into its web address, and passive DNS databases kept the addresses. Meta suspended sharing on October 20 and has not said what happens to existing links.
- 04Laminar's flow-1 Costs $0.0011 a Trace, but Claude Opus 5 Scores Higher on Its Own BenchmarkThe Y Combinator S24 company says its reinforcement-learned error finder matches GPT-6-sol at one-twentieth of the cost. Its launch post and its benchmark table put the multiple at 20 and 23, and show a higher score for Opus 5.