NVIDIA releases Nemotron 3 Diarization, an open model that tells who spoke when

NVIDIA published its Nemotron 3 Diarization model on Hugging Face on September 23, 2026. The model works out who spoke when in an audio recording and is designed for meetings, customer calls and podcasts. The weights are open for anyone to download, and the license allows commercial use.

What speaker diarization is

When several people speak in a recording, speech recognition models write down the words but cannot tell on their own who said them. A diarization model fills that gap: it listens to the recording and marks the seconds in which each speaker talks. The model does not write down the words and does not identify people; it gives speakers anonymous labels such as "speaker 1" and "speaker 2". Combined with a speech recognition model, it produces a transcript that shows who said what.

What it can do

According to the model card, it can separate up to eight speakers in the same recording; the previous version was limited to four. It also handles overlapping speech. It can process a recording afterwards or run on live audio, where the delay can be brought below one second. It is a small model of about 100 million parameters. NVIDIA's list of supported hardware includes GeForce cards from the RTX 3050 to the RTX 5090, so the model can run on your own server or a capable desktop computer. It can be used with NVIDIA's NeMo library, Hugging Face Transformers and a lightweight C++ runtime.

Who can use it

The model is released under the OpenMDW 1.1 license. According to the model card, it is ready for commercial and non-commercial use. Anyone who redistributes it must keep the license text and copyright notices.

According to NVIDIA, the model ranks first on the VoiceArena diarization leaderboard, which uses English conversations, and makes fewer errors than its predecessor on all eight test sets. The company also shares that on a set of two-person phone calls the new model trails the old one slightly. The model card says that in recordings with more than eight speakers some speech can be missed or assigned to the wrong person, and that accuracy can drop on very long recordings and in noisy or echoing rooms. The training data covers English, Mandarin, Hindi and other languages; Turkish is not mentioned separately. Before trying it on your own recordings, it is worth following evaluations from independent researchers as well.

Sources: Hugging Face: NVIDIA Nemotron 3 Diarization announcement, Hugging Face: nvidia/Nemotron-3-Diarization.