Nvidia ships open-weight Nemotron 3.5 Transcribe and an updated Granary speech dataset

Nvidia has released Nemotron 3.5 Transcribe, a free, open-weight model for speech transcription covering as many as 40 languages, together with the second version of Granary, its open speech dataset. According to TechCrunch, which reported the release on October 28, the two releases target developers building applications for business meetings and medical documentation, two areas where AI tools have already gained firm traction in the workplace.

What Nvidia shipped

Transcribe is a 30-billion-parameter model designed to run on one chip, according to TechCrunch. That footprint spans Nvidia's hardware range, from the Jetson AGX Thor robotics controller and the DGX Spark chip found in new laptops up to Blackwell datacenter GPUs. The model joins an existing lineup of Nvidia open-weight voice models, including Parakeet, a much smaller offering at 1.1 billion parameters.

Granary, Nvidia's open speech dataset, was refreshed to version 2 at the same time. Nvidia says developers can combine the model, the updated dataset, and its NeMo Speech tooling to tailor applications to the vocabulary of a specific industry in each supported language, an important consideration for fields like medicine where general-purpose models often stumble on specialized terms.

Capabilities and limits

According to TechCrunch, Nemotron 3.5 Transcribe can translate between English and 31 languages, process recordings as long as an hour, and separate as many as eight speakers in a single file. Nvidia claims it outperforms competing models on benchmarks, including Hugging Face's Open ASR Leaderboard.

The notable gap is streaming. The model works only on recorded audio rather than live streams, which rules it out, at least in its current form, for real-time voice agents and live captioning, applications where low latency is the whole point. Developers building interactive voice products would need a separate component for that job.

A crowded stretch for voice AI

The release lands during a run of voice-focused launches. As TechCrunch notes, Mistral shipped Voxtral 3 earlier in October, ElevenLabs introduced Realtime for instantaneous speech in voice agents, and Salesforce released its own voice model, Launa-Omni. Nvidia's angle is making Transcribe free and open-weight, giving developers a model they can download, study, and fine-tune on their own infrastructure instead of relying solely on a hosted API.

Why it matters

Transcription is one of the few AI categories with proven, everyday demand, and meeting notes plus medical records are among its most reliable revenue sources. An open-weight, multilingual option that fits on a single chip lowers the barrier for smaller teams to build competitive products, and it opens the possibility of on-device or on-premise deployments, which matters for privacy-sensitive domains such as healthcare.

The paired release of Granary 2 is arguably as significant as the model itself. Customization for industry jargon and less common languages depends on the availability of good training data, and an openly licensed dataset gives developers raw material they would otherwise have to buy or build themselves.

The limits frame the opportunity, though. Without streaming support, Nemotron 3.5 Transcribe competes in batch transcription and post-processing workflows rather than the fast-growing market for live voice agents, where rivals like ElevenLabs are concentrating. For developers already working with recorded audio, however, a free model claiming benchmark-leading accuracy that runs anywhere from a laptop chip to a datacenter GPU is a practical release with few direct drawbacks beyond the streaming question.