Nemotron 3 Diarization

Released on 23 September 2026, Nemotron 3 Diarization is NVIDIA's open-weight model for telling speakers apart in audio. Diarization is the step in a transcription pipeline that works out who spoke when. The model has about 100 million parameters, small enough to run alongside a speech-recognition model.

What Nemotron 3 Diarization offers

• Speaker probabilities: every 10 milliseconds, gives the probability that each of up to eight speakers is talking, doubling the four-speaker limit of NVIDIA's earlier Streaming Sortformer.
• Offline or streaming: works offline on full recordings, or live with latency settings of 1.04, 0.64 or 0.32 seconds.
• Audio input: takes 16 kHz mono audio.
• Hosted routes: NVIDIA names Baseten, DigitalOcean and Argmax as hosted options.

Clients

Streaming suits live meeting and call-centre products, and the offline mode suits recorded audio.

Key facts

• Licence: OpenMDW 1.1, which allows commercial use.
• Languages: English, Mandarin, Hindi, Kannada, Telugu and Bengali on the model card. NVIDIA says training covered 21 languages and evaluation was in English.
• Benchmark: first of twelve systems on Voice Arena's Diarization-Bench, at a 14.72 per cent error rate, labelled early results and read on 25 September 2026.
• Developer's figures: 12.73 per cent on DIHARD III, per NVIDIA.