NVIDIA released Nemotron 3 Diarization, a new 100M-parameter open-weight speaker diarization model available on Hugging Face. The model is designed to answer who spoke and when during conversations, tracking up to eight speakers simultaneously including during moments of overlapping voices. Released under the OpenMDW License 1.1 for commercial use, the model supports both offline recordings and real-time streaming on Linux systems using NVIDIA GPUs.
The architecture processes 16 kHz audio into Mel-spectrogram features, which are analyzed by a 31-layer Transformer encoder utilizing rotary positional embeddings. Using an Arrival-Order Speaker Cache and a FIFO queue, the model maintains stable speaker labels across streaming chunks without requiring continuous re-matching. In initial evaluations, Nemotron 3 Diarization ranked first on Voice Arena's Diarization-Bench, achieving a 14.72% diarization error rate and showing significant performance gains across multiple benchmark conditions compared to previous streaming baselines.
This release matters because accurate speaker diarization is essential for producing reliable speaker-attributed transcripts used in meeting tools, call analytics, and voice-agent memory applications.