NVIDIA Releases Nemotron 3.5 Transcribe, Granary v2

NVIDIA has released Nemotron 3.5 Transcribe and Granary v2, a pair of multilingual transcription models built to give developers more control over speech-to-text workflows. The models can be guided with prompts, which means users can supply context, terminology, or formatting instructions instead of relying only on a generic transcription pass. That is a meaningful shift for applications where accuracy depends on domain knowledge, such as legal, medical, technical, or multilingual customer-support recordings. Promptable transcription can help a model handle names, acronyms, product terms, and preferred output styles. For example, a team working with specialized audio can tell the model what kind of content to expect or how a transcript should be structured. The goal is not just word-for-word conversion, but a more useful text output that fits downstream translation, search, summarization, or analysis. NVIDIA says the release is aimed at improving accuracy and control across languages for developers building transcription, translation, and audio-analysis applications. Granary v2 and Nemotron 3.5 Transcribe add to a growing ecosystem of open and accessible speech models. That matters because many teams want to fine-tune, self-host, or integrate transcription into their own pipelines rather than depend entirely on closed APIs. At the same time, prompt-guided systems introduce new evaluation questions: teams need to check whether prompts improve consistency, how well the models handle accents and noisy audio, and whether performance holds across languages. The practical opportunity is in specialized workflows. A generic transcript is often only the starting point; the real value comes when the output is clean, terminologically correct, and easy to process automatically. NVIDIA's update signals that competition in speech AI is moving toward controllability and domain fit, not just raw transcription accuracy. Developers will want to benchmark the models on their own audio before deciding where they fit.