Nemotron 3 Diarization - Who Said That?

Speaker diarization answers the question of who spoke when in a recording. In this 13-minute video, Sam Witteveen looks at NVIDIA's Nemotron 3 Diarization, a new model that handles both batch processing and streaming. The description links the Hugging Face blog post and the model card.

The model is introduced at 1:27, and at 2:30 he places it within the Nemotron Speech family. A short definition of diarization follows at 4:09. At 4:26 he connects it to agents and voice, then at 5:41 explains how these models are scored with diarization error rate (DER). The description says the model beats many others at telling speakers apart. The chapter list has no separate benchmark section, so check that claim against the DER figures on the model card.

From 6:26 he breaks down how Nemotron 3 Diarization works. The demo starts at 6:58 on a DGX Spark with NVIDIA NeMo and runs close to four minutes. At 10:52 he processes a full podcast. The final chapter, from 11:38, exports the result as plain text or as an SRT subtitle file.

The model card fills in the specs. It is a 100M-parameter Transformer encoder that handles up to eight speakers. A single checkpoint covers every latency setting, from streaming to offline. Training data spans English, Mandarin, Hindi, Kannada, Telugu and Bengali, and the license is the OpenMDW License Agreement 1.1, which the card describes as ready for commercial or non-commercial use.