Model Release and Capabilities
Nvidia released Nemotron 3.5 Transcribe, a free, open-weight model that transcribes audio in up to 40 languages, alongside version 2 of its Granary open-source speech dataset.
The model translates 31 languages to and from English, processes recordings up to an hour long, differentiates up to eight speakers, and performs competitively on benchmarks such as Hugging Face’s Open ASR Leaderboard.
Target Applications and Hardware
These tools target developers building applications for business meetings and medical transcription. Developers can use the model, dataset, and NeMo Speech tools to customize vocabulary for specific industries. The current version processes recorded audio rather than live streams.
Containing 30 billion parameters, the model runs on hardware ranging from the Jetson AGX Thor robotics controller and DGX Spark chips to Blackwell data center hardware, joining Nvidia's existing open-weight voice models like Parakeet.
Industry Context
The release arrives amid a surge of voice AI developments, following Mistral's release of Voxtral 3, ElevenLabs' launch of Realtime for voice agents, and Salesforce's introduction of the Launa-Omni voice model.
