Sakhanda Wire
NVDA $154.00 -1.37% MSFT $506.28 -2.26% GOOGL $280.87 -1.00% META $620.34 -0.97% AMZN $256.88 -0.34%
← Back to the news

NVIDIA releases open-weight model for tracking up to eight speakers

NVIDIA has released Nemotron 3 Diarization, a 100M-parameter open-weight model available on Hugging Face. It identifies when each participant speaks and can track up to eight speakers, including overlapping voices. The model uses one checkpoint for recorded audio and real-time streaming, and its OpenMDW License 1.1 permits commercial use.

The system processes 16 kHz single-channel audio and produces time intervals with anonymous speaker labels. It runs on Linux through NVIDIA NeMo with Ampere, Ada Lovelace, Hopper or Blackwell GPUs. Recommended input-buffer latency ranges from 30.4 seconds for offline-style processing to 0.32 seconds for ultra-low-latency streaming. These figures exclude compute, networking and ASR time.

In Voice Arena’s initial Diarization-Bench, Nemotron 3 Diarization ranked first among 12 systems and 17 configurations, recording 14.72% DER on 139 English conversations. NVIDIA reported a 41.0% mean relative DER reduction versus its four-speaker Streaming Sortformer baseline at 1.04 seconds. Performance can deteriorate with more than eight speakers, heavy noise, reverberation or far-field recordings. The model is not yet available through Hugging Face Inference Providers.

Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.

Originally published by MarkTechPost on

Read the original on MarkTechPost ↗

The reporting and the image belong to MarkTechPost and are credited and linked above; the text on this page is Sakhanda Wire's own summary.

← Back to the news

More stories

All the latest news