Nvidia Releases Weights for Audio Speaker Diarization Model

On September 23, 2026, Nvidia released the weights of the Nemotron 3 Diarization model under the OpenMDW-1.1 license. The model determines "who spoke when" in audio with up to 8 speakers and ranks first out of 12 systems on the VoiceArena Diarization-Bench with a DER of 14.72%.

Compiled by Sergey KostenchukPublished 2026-09-24Updated 2026-09-24

2026-09-24 Research NVIDIA

🌟 Nvidia Releases Weights for Audio Speaker Diarization Model

On September 23, Nvidia published the weights of the Nemotron 3 Diarization model (~100 million parameters) under the OpenMDW-1.1 license, which permits commercial use. The model determines "who spoke when" for up to 8 speakers — twice as many as its predecessor, Streaming Sortformer. On VoiceArena, it ranks first out of 12 systems with a DER of 14.72%. Nvidia claims a 41% reduction in DER and an increase in RTFx to 865×.

🌍 ASR provides the words, while this model provides speaker labels: meeting minutes with names, call analytics, and memory for voice agents. Open weights remove the barrier to embedding "who said what" into products.

👤 The weights are available on Hugging Face, with deployment via NeMo 3.0 or NeMo-Speech.cpp, and a demo Space is available. The model is anonymous: speaker_1…8 without identity, and the recording length is unlimited.

Source 1: https://huggingface.co/blog/nvidia/nemotron-diarization Source 2: https://huggingface.co/nvidia/Nemotron-3-Diarization

Sources