← All news·2026-09-27·3 min read

Nvidia has open sourced its model for distinguishing voices during conversations

Nvidia has open-sourced their Nemotron 3 Diarization model, which can identify up to 8 different speakers and is currently the most accurate diarization model available.

aiaudio

Nvidia has made their Nemotron 3 Diarization model available on Hugging Face for everyone to use. This model takes audio input and determines which speaker is talking at each point in time. It supports up to 8 simultaneous speakers and even works when multiple people are speaking at the same time. Currently, this model is the most accurate on the VoiceArena Diarization Benchmark v1, with an error rate of 14.72% compared to the next best model’s 19.3%. Additionally, this model is an improvement over Nvidia’s last model, reducing the average error rate by 41%.

I think it’s interesting that this is a specialized task rather than something a chatbot would need to do. With only 100 million parameters, this model doesn’t seem to be sacrificing much quality like small general purpose models often do. You can find the weights on Hugging Face, so you can download the model and use it yourself without paying a monthly fee or calling someone’s API. This model is small enough to run on a typical server instead of requiring tens of thousands of dollars worth of hardware. When combined with a speech recognition model like Parakeet, you could build a voice assistant that correctly identifies which person said what.

Source: the-decoder.com

Free course

Stop reading about AI — start building with it

The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.

Start free →
EAEvgenii Arsentev

Author

Evgenii Arsentev

PhD · Founder & CEO, Dad4Kids