Nemotron 3 Diarization

Nemotron 3 Diarization is an open-weight speaker-diarization model from NVIDIA, released on 23 September 2026 with a model card, a Hugging Face blog post and coverage in MarkTechPost the same day. Diarization is the part of a transcription pipeline that works out who spoke when: the model takes 16 kHz mono audio and returns, every 10 milliseconds, the probability that each of up to eight speakers is talking — twice the four speakers of NVIDIA’s earlier Streaming Sortformer. It runs offline on whole recordings or in streaming mode with latency settings of 1.04, 0.64 and 0.32 seconds, which is what a live meeting or call-centre product needs. It has about 100 million parameters, small enough to run next to a speech-recognition model rather than instead of one. The model card lists English, Mandarin, Hindi, Kannada, Telugu and Bengali, while NVIDIA’s blog says the training data covered 21 languages and that evaluation was in English, so non-English performance should be tested before relying on it. The licence, OpenMDW 1.1, allows commercial use; NVIDIA names Baseten, DigitalOcean and Argmax as hosted routes, with no per-use price of its own. On Voice Arena’s independent Diarization-Bench — 139 far-field English sessions, about 22 hours, labelled “early results” — it ranked first of twelve systems with a 14.72 per cent diarization error rate when read on 25 September 2026, ahead of DiariZen WavLM Large at 19.34 and pyannoteAI Precision-3 at 20.56. NVIDIA’s own figures, including 12.73 per cent on DIHARD III and a slight regression on two-speaker CALLHOME, are its own measurements and are not stored as scores.