Nemotron 3 Diarization: find who spoke when and every interruption

Nemotron 3 Diarization is NVIDIA's open 100M-parameter model that works out who is talking, and when, for up to eight speakers. You give it a recording and it gives you time ranges such as "0.40 to 2.10 seconds, speaker_0". Where two ranges from different speakers overlap, someone cut in. It does not write the words, so you pair it with a speech-to-text model and line the two up on one timeline. You can try it free in the browser first. Running it yourself needs Linux and an NVIDIA GPU. I have not run the checkpoint myself; the commands below come from NVIDIA's model card and blog post.