Breaking

NVIDIA Releases Nemotron 3 Diarization for Overlapping Multi-Speaker Audio

September 23, 2026 at 11:03 EDT

Foundation Models
Open Source

NVIDIA has released Nemotron 3 Diarization, an open-weight 100M-parameter speech model that tracks who spoke when in live or recorded conversations, including overlapping speech. The model supports up to eight speakers and is available through Hugging Face for developers building speaker-aware transcripts, meeting tools, call-center analytics and voice-agent memory systems.

September 23, 2026 · NVIDIA

Who spoke when, even in overlap

Nemotron 3 Diarization is a 100M-parameter model for live and recorded conversations, assigning anonymous speaker timelines across as many as eight speakers.

100M

model parameters

speaker channels

14.72%

VoiceArena DER

0.32s

lowest recommended buffer

Speaker capacity doubles from the prior public Sortformer line

4Streaming Sortformer

8Nemotron 3 Diarization

The new model expands anonymous speaker channels from four to eight while keeping support for overlapping intervals.

Why developers care

Speaker timelines make meetings, interviews, support calls and voice-agent memory more useful than plain transcripts.

What still limits it

The preview license is for internal testing and evaluation, and the lowest buffer figure excludes full application latency.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.