Breaking

NVIDIA Releases Nemotron 3.5 Transcribe for 40-Language Audio Workflows

October 29, 2026 at 02:51 EDT

Foundation Models
Open Source

NVIDIA released Nemotron 3.5 Transcribe on Oct. 28 as an open-weight model for recorded speech, combining transcription, speaker attribution, translation and audio question answering. It transcribes 40 languages, handles recordings up to one hour and has about 4.3 billion active parameters in a 34-billion-parameter model. NVIDIA's announcement and model card describe the release.

October 28, 2026 · NVIDIA

One model for recorded speech

Nemotron 3.5 Transcribe combines multilingual transcription, speaker attribution, translation and audio questions in an open-weight model.

English open-weight accuracy

Mean word error rate across nine public test sets · lower is better

4.71%

Nemotron 3.5 Transcribe

4.94%

Qwen3-ASR-1.7B

Each stacked block represents 0.5 percentage points; the top block is partial.

transcription languages

4.3B

active parameters

1 hour

maximum recording length

What it brings

Promptable transcript styles, contextual vocabulary, speaker labels and translation in one model.

Where it stops

Designed for complete recordings, not live streaming; NVIDIA's fastest serving optimizations were outside the launch-day open-source stack.

Continue reading

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