NVIDIA has released Nemotron 3.5 Transcribe, an open-weight speech model that can transcribe, translate and label speakers in one checkpoint. It supports recordings up to an hour and runs on a single NVIDIA GPU, giving developers a model they can adapt and run where they choose.
NVIDIA Watch analysis
What happened
The model transcribes 40 languages, translates 31 to or from English, and supports speaker attribution in 13 languages, with up to eight speakers labelled. Developers can guide its output with natural-language prompts, including requests for domain vocabulary or a particular transcript format. It is designed for recorded audio, not live streaming, and supports word-level timestamps and audio questions, though the article says audio question answering is trained in English.
NVIDIA says Nemotron 3.5 Transcribe ranks first among open-weight models on the Hugging Face Open ASR Leaderboard’s English track, with an average word error rate of 5.33%. It also reports leading results on multilingual transcription and speaker-attribution benchmarks. These are benchmark results cited in NVIDIA’s announcement, not a guarantee that every language, recording or speaker will perform equally well.
Alongside the model, NVIDIA has released Granary v2, a speech dataset collection and reusable curation pipeline. Its public manifest covers 463,062 hours across 48 languages and 16 source corpora. NVIDIA says fine-tuning on medical speech reduced word error rate by 45.8% relative to the best prompt in a held-out-speaker case study. That is a result from a specific adaptation study, not a general improvement figure for every use.
Why it matters
Speech tools often ask developers to stitch together transcription, translation and speaker labelling, then wrestle the output into a format that suits the job. Combining those tasks with prompt-level control could make it easier to build practical workflows for meetings, classes or specialist recordings. The option to run the model locally or on a chosen GPU also gives developers more control over where audio is processed.
Granary v2 matters because adaptation depends on training data with clear targets. Its manifests and reusable curation pipeline give developers material and tools to explore how transcript conventions can be changed, rather than treating every domain-specific error as a prompting problem.
Our read
This is a substantial open-weight speech release, not merely another model name on a leaderboard. The useful combination is broad language support, speaker labels and adaptable transcripts, backed by a public data and curation effort. Developers should still test it against their own recordings: names, overlapping speech and specialist vocabulary remain exactly the sort of details that can make a polished transcript look more certain than it is.
What to watch
- How the model performs on independent tests and on languages or accents important to real deployments.
- Whether local-serving tools make deployment straightforward beyond NVIDIA’s cited hardware.
- How developers use Granary v2 and fine-tuning to improve specialised transcription.
- Whether future releases add support for live speech.
Discussion spark: For speech applications, would you prioritise one adaptable model for transcription, translation and speaker labels, or separate specialist tools that can be tested task by task?
Sources and evidence
- NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt (28 October 2026, 14:58 UTC)
Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.