AI & Machine Learning

NVIDIA opens up Nemotron 3.5 Transcribe. Meeting and medical teams get a free model, but not a live one

2 min read

Stained-glass style illustration of three coloured speech bubbles sending sound waves through an open book and coming out as colour-coded lines of text, one per speaker, for NVIDIA's open-weight Nemotron 3.5 Transcribe speech model.

NVIDIA has released Nemotron 3.5 Transcribe, a free open-weight speech model, with the Granary v2 dataset. It suits recorded meetings and dictation, but speaker labels cover only 13 languages and live audio is out of scope.

In brief

On 28 October 2026 NVIDIA released Nemotron 3.5 Transcribe, an open-weight model that combines transcription, translation, speaker labelling and audio question answering in one checkpoint. It came with Granary v2, an open speech dataset and curation pipeline. The weights are published under the OpenMDW-1.1 licence, which grants free use and modification, including commercial use, provided licence and copyright notices are kept.

What NVIDIA says it does

  • Languages. Transcription in 40 languages; translation of 31 of them to or from English.
  • Speakers. Labels up to eight speakers, but speaker attribution works in only 13 of the 40 languages.
  • Length. Recordings up to about an hour. NVIDIA says it is built for recorded audio, not live streaming.
  • Accuracy. NVIDIA states the model ranks first among open-weight models on the English track of the Hugging Face Open ASR Leaderboard, at 5.33% average word error rate (WER), with API models excluded.
  • Size. About 30 billion parameters, roughly 4 billion active per token, running on a single NVIDIA GPU.

What it means for meeting and medical transcription

For teams building meeting tools, one self-hosted model that transcribes, labels speakers and can answer questions about the recording, such as what was decided (English only), removes a pipeline of separate services, and keeps audio inside your own infrastructure. The 13-language limit on speaker labels matters for multinational meetings: check your languages against the model card's task matrix first.

For clinical dictation, NVIDIA's own case study on an Indian medical speech benchmark reports fine-tuning cut held-out WER from 22.44% for the best-prompt baseline to 12.17%. That is a vendor result on one dataset, and NVIDIA notes a specialist model still did better on brand-name drugs.

What it cannot do yet

There is no live streaming, so real-time captioning and voice agents need another model. Audio question answering is trained in English only. Benchmarks are vendor-reported; test on your own recordings, with your accents and vocabulary, before relying on it.

Sources and scope

Capabilities, figures and the case study are from NVIDIA's Hugging Face announcement (28 October 2026). Licence terms are from the model card and LICENSE file, checked 28 October 2026. TechCrunch's report is secondary corroboration. Leaderboard results are NVIDIA claims, not independently tested. Implications for buyers are ByteNib editorial analysis.

Drafted with AI assistance and reviewed by the editor before publication.

Continue exploring: AI & Machine Learning analysis, the related implementation tutorial, and the structured learning path.