Nvidia's free Nemotron 3 Diarization model separates 8 speakers

Nvidia has built a broad catalogue of freely downloadable AI models, and this release extends that strategy into the unglamorous but commercially valuable job of working out who said what in a conversation. The model is small by the standards of modern AI (about 100 million parameters) and is designed to run on audio as it arrives, rather than only on files after the fact.

The company is not just claiming the model works; it is pointing to an external scoreboard. In the VoiceArena Diarization Benchmark v1, the freely available Nemotron 3 leads with a 14.7% DER and outperforms its predecessor, Streaming Sortformer, by 41%. On the same family of tests, the model sits first with a 14.72 percent error rate, ahead of the next best system at 19.3 percent. The benchmark is described as strict: overlapping speech counts, and even tiny misalignments at speaker transitions are scored as errors. Across eight test scenarios, the new model cuts the error rate by an average of 41 percent compared with Streaming Sortformer when using a 1.04-second buffer.

The practical caveat is that pushing for lower latency costs accuracy. The audio buffer can be set to four levels ranging from 30.4 down to 0.32 seconds, and shorter buffers generally reduce accuracy. The real test is whether teams that build meeting notes, call summaries, or accessibility tools find the free model good enough outside the lab, where reverb and crosstalk are common. If it holds up, the pricing pressure would fall on paid transcription and diarization services first, while Nvidia gains another reason for developers to stay inside its ecosystem.