Nvidia’s New Transcription Model Brings Multilingual Speech AI to One Chip

Nvidia has released Nemotron 3.5 Transcribe, a free, open-weight model designed to turn recorded speech into text across up to 40 languages. The company also released version 2 of Granary, its open source speech dataset, giving developers another resource for building and refining speech applications.

The release targets practical tools rather than a single showcase. Nvidia points to business meeting apps and medical transcription as two areas where developers could use the model. Nemotron 3.5 Transcribe can also translate 31 languages to and from English, handle recordings up to an hour long, and separate up to eight speakers in one recording.

A speech model built for longer, messier recordings

Speaker separation matters when several people talk during the same meeting, interview, or medical discussion. Instead of producing one block of text, the system can tell up to eight speakers apart, helping an app organize a conversation around who said what.

Nvidia says Nemotron 3.5 Transcribe beats rival systems on benchmarks, including Hugging Face’s Open ASR Leaderboard. Those performance claims place the model in a crowded field where developers are comparing speech systems not only by accuracy, but also by language coverage, speaker handling, and the hardware needed to run them.

There is one clear limit: Transcribe works with recorded audio, not livestreams. That distinction matters for apps that need instant captions or live voice processing, since this release is aimed at audio that already exists rather than speech arriving in real time.

Still, Nvidia says the model can run on a single chip despite having 30 billion parameters. It can run on the Jetson AGX Thor robotics controller, the DGX Spark chip powering new laptops, and Blackwell datacenter chips. That range gives developers a way to consider the same model across robotics, laptops, and datacenter systems.

Granary gives developers more room to customize

Nvidia says developers can combine Nemotron 3.5 Transcribe with the updated Granary dataset and tools such as its NeMo Speech model. The goal is to customize apps so they recognize an industry’s jargon in each language, which could matter when ordinary speech systems struggle with technical vocabulary.

That open approach also gives developers more control over how their speech tools work. They can use the model and dataset as starting points for apps designed around meetings, healthcare, or other specialized settings, rather than relying on a fixed transcription service.

Nemotron 3.5 Transcribe joins Nvidia’s other open-weight voice models, including Parakeet, which has 1.1 billion parameters. The contrast between Parakeet’s smaller size and Transcribe’s 30 billion parameters shows the range Nvidia is building into its voice lineup, from a compact model to a much larger system aimed at broader transcription tasks.

A busy market for open and voice models

Nvidia’s new voice entry arrives during a busy period for AI speech releases. Mistral shipped Voxtral 3 earlier this month, ElevenLabs launched Realtime for instant speech in voice agents, and Salesforce released Launa-Omni, its voice model.

Recent projects also show how open-source and open-weight systems can be refined for frontier-level performance and stronger safety. AI security company Alice worked with AI21 Labs and Inflection AI on projects involving Nvidia Nemotron models. With AI21, Alice post-trained Nemotron 3.5 Super VL to protect enterprise agents against indirect prompt injection. With Inflection, it helped develop red teaming datasets to reduce sycophancy in Nemotron 3 Super.

That work adds another layer to the open-model debate. Releasing model weights gives developers more freedom to adapt a system, but the final quality and safety can depend on the data, testing, and post-training applied afterward.

Jensen Huang, Nvidia’s president and CEO, has also used Nvidia GTC Berlin 2026 to argue that open models can beat the frontier. Nemotron 3.5 Transcribe fits that argument by pairing open weights with broad language support, speaker separation, translation, and the ability to run on a single chip.

For developers, the appeal is straightforward: one model can cover many languages, recordings can run for up to an hour, and specialized apps can be shaped around industry terms. The tradeoff is just as clear. Anyone building live transcription still needs a different system because Transcribe does not handle livestreams.

Stay connected via Google News

Follow us for the latest AI News