Streaming Nemotron 3 Diarization
retained text
I’ve been playing with [Nemotron 3 Diarization](https://huggingface.co/nvidia/Nemotron-3-Diarization), and it fills a gap I’ve had with local voice agents: keeping track of *who* is speaking. It’s a diarization model, so it gives you speaker labels rather than transcriptions or people’s names. It can stream its output and track up to eight speakers. I’ve been trying it with one-second streaming chunks, and the quality has been really good in my tests. I plugged it into my [speech-to-speech](https://github.com/huggingface/speech-to-speech) setup on a DGX Spark and connected it to a Reachy Mini. The fun part is watching someone new speak, then seeing the robot ask their name and remember it for the conversation (that's the video!). It has day-zero Transformers integration and is getting a commercial-friendly license.
document fetched 2026-09-24 00:45 · canonical https://reddit.com/r/LocalLLaMA/comments/1wo8tr5/streaming_nemotron_3_diarization