Nvidia released Nemotron 3.5 Transcribe on Wednesday, a free, open-weight speech model that folds transcription, speaker labels, translation, and spoken question answering into a single promptable checkpoint. It transcribes audio in up to 40 languages, translates 31 of them to and from English, and can tell up to eight speakers apart in recordings up to an hour long. Nvidia says it sits first among open-weight models on the Hugging Face Open ASR Leaderboard for English, and it shipped alongside version 2 of Granary, the company's open speech dataset and curation pipeline, whose public manifest covers 463,062 hours across 48 languages.
The engineering choice that matters is sparsity. Transcribe is a 30-billion-parameter mixture-of-experts model that activates roughly 4 billion parameters per token, pairing a speech encoder with a hybrid Mamba-Transformer language model. That is why Nvidia can credibly claim it runs on a single chip, from a Jetson AGX Thor edge controller or a DGX Spark desktop up to Blackwell data-center hardware. In the company's own benchmark, one RTX PRO 6000 Blackwell Server Edition GPU processed 100 hours of audio in under three minutes. Builders can also steer the output with plain-language prompts and pass in domain vocabulary, so a medical or legal transcript can spell its own jargon correctly instead of guessing.
The limits are just as clear, and they define where this model fits. It works on recorded audio, not livestreams, so it will not power live captions or a real-time voice agent by itself. That is not a flaw so much as a lane: meetings, dictation, call archives, and media libraries are batch jobs, and they are also where transcription spending actually lives today. Nvidia is releasing the weights under the OpenMDW-1.1 license for commercial or non-commercial use, with hosted endpoints already offered by providers including Baseten and DeepInfra for teams that do not want to run a GPU themselves.
The timing is not accidental. Voice has become the busiest corner of the release calendar: Mistral shipped Voxtral 3 earlier this month, ElevenLabs launched Realtime for instant speech in voice agents, and Salesforce released its own Launa-Omni voice model. Against that field, Nvidia is not selling a hosted voice service so much as it is selling the case for owning the model. If your transcripts contain patient details, client calls, or unreleased product discussions, running an open-weight model on your own hardware changes the privacy and cost conversation entirely — no per-minute API bill, and no audio leaving the building.
For practitioners, the practical test is simple. Take an hour of your hardest real audio — overlapping speakers, accents, domain terms — and run Transcribe against whatever you pay for now. Compare word error, speaker labels, and total cost, including the GPU time. If an open model that activates 4 billion parameters ties or beats a closed service on your own data, the default for speech-to-text quietly flips from renting to running. That is the shift Nvidia is betting Granary v2 will accelerate: not just a model release, but the data and recipes for every team to adapt it to its own voice.
