Analysis
NVIDIA introduced Nemotron 3.5 Transcribe and Granary v2 in an October 28 article on Hugging Face. The open-weight model combines transcription, speaker attribution, translation and audio questions in one checkpoint. NVIDIA lists transcription in 40 languages, translation to or from English for 31, and speaker attribution in 13. Those differing totals matter: support for a language in one task should not be assumed to cover every other function.
The article pairs prompting with training data and adaptation recipes. Developers can specify transcript conventions, while Granary v2 provides manifests and a reusable curation pipeline. NVIDIA’s medical adaptation case study reports a 45.8% relative reduction in word error rate against its best-prompt baseline on held-out speakers. That comparison is more informative than a result that mixes the examples used for tuning with those reserved for evaluation.
A useful application test begins by deciding what counts as a correct transcript. A faithful record might preserve hesitation; a searchable document might standardize numbers and terminology. Speaker changes and specialist names require their own checks. These choices affect what a model is asked to produce and what an evaluator counts as an error. Aggregate accuracy alone cannot settle which written form fits a particular workflow.
The availability details also deserve attention. NVIDIA says the model handles recorded audio rather than live streaming, with quantized versions and some local deployment support still forthcoming. We have not run the checkpoint or reproduced its benchmarks. Teams assessing it should compare output against recordings from their intended setting, with separate checks for words, speakers and answers, before trusting a polished transcript as the record of what occurred.
Source: NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt · 2026-10-28
Why it matters
The right transcript depends on task, language and formatting conventions; deployment and evaluation should reflect those distinctions.