Updated

Nemotron 3 Diarization

NVIDIA released Nemotron 3 Diarization, a 100-million-parameter open-weight model that identifies who spoke when in real time for up to eight speakers, independent of speech recognition. NVIDIA reports it ranks first in VoiceArena's initial Diarization-Bench with a 14.72 percent diarization error rate.

By No Code MBA · Facts checked

Provider
NVIDIA
Announcement date
Status at last check
Generally available
Model family
Nemotron 3
Preceding model
Streaming Sortformer (diar_streaming_sortformer_4spk-v2.1)
Capabilities
Audio

What changed

Nemotron 3 Diarization is a roughly 100M-parameter streaming model that labels who spoke when for up to eight speakers, double the four supported by NVIDIA's previous Streaming Sortformer checkpoint. It takes 16 kHz mono audio with buffer latencies as low as 0.32 seconds. NVIDIA reports a 14.72% diarization error rate on VoiceArena's Diarization-Bench v1, ranking first, and a 41% mean relative error reduction versus Sortformer across eight datasets.

Access and availability

Open weights on Hugging Face at nvidia/Nemotron-3-Diarization under the OpenMDW License 1.1, designed for Linux with NVIDIA Ampere, Hopper or Blackwell GPUs. NVIDIA lists Baseten and DigitalOcean for hosted deployment and Argmax Pro SDK 3 for on-device use.

Official access documentation

What you can build

Download the open weights and pair its speaker timestamps with a transcription model to build meeting notes or call summaries without sending audio to a cloud API.

Limitations

It does not transcribe words and must be paired with a speech recognition model for speaker-attributed transcripts. More than eight speakers, noise, reverberation, far-field audio and very long recordings can raise errors, and the published throughput figures are batched, not single-stream latency.

Sources and release history