Analysis
A customer pauses, corrects a product name, then gives the order number. Should the transcript keep the hesitation, normalize the number, or get the name right above all?
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. But NVIDIA lists different language coverage by task: transcription in 40 languages, translation to or from English for 31, and speaker attribution in 13. Those are separate capabilities, not a blanket promise that every function works across every listed language.
The release pairs prompting with training data and adaptation recipes. Developers can specify transcript conventions; Granary v2 supplies manifests and a reusable curation pipeline. In NVIDIA’s medical adaptation case study, the adapted model cut word error rate by 45.8% relative to its best-prompt baseline on held-out speakers. That comparison is useful because the evaluation speakers were kept apart from those used for tuning. It remains NVIDIA’s result, not an independently reproduced benchmark.
A transcript can be a faithful record, a searchable document, or an input to another workflow. Those jobs may disagree about whether to preserve hesitation or standardize numbers and terminology. Speaker changes and specialist names need their own checks; a clean aggregate score cannot decide which version your staff can rely on.
NVIDIA says the model handles recorded audio, not live streaming. Quantized versions and some local deployment support were still forthcoming at announcement. We have not run the checkpoint or tested it in a deployment.
Before adopting it, write down what “correct” means for your workflow, then compare its words, speaker labels and answers against recordings from that setting. The transcript is ready only when someone downstream can trust the record.
Source: NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt · 2026-10-28
Why it matters
Operators should define and test the transcript their workflow needs before treating a polished output as the record of what happened.
