Nvidia has launched Nemotron 3.5 Transcribe, a free, open-weight model designed to transcribe audio in up to 40 languages. The company also introduced version 2 of its open-source speech dataset, Granary. Nemotron 3.5 can translate 31 languages to and from English, handle recordings up to an hour long, and distinguish between up to eight speakers. Nvidia claims the model outperforms rivals on benchmarks, including Hugging Face’s Open ASR Leaderboard.
The tools are aimed at developers building applications for business meetings and medical transcription, two areas where AI has already made significant inroads. Nvidia emphasizes that developers can customize apps to recognize industry-specific jargon in each language using the model, the updated Granary dataset, and tools like its NeMo Speech model. However, a notable limitation is that Transcribe works only with recorded audio, not live streams.
Despite this, Nvidia highlights that Transcribe, with 30 billion parameters, can run on a single chip, including the Jetson AGX Thor robotics controller, DGX Spark chip, and Blackwell datacenter chips. It joins Nvidia’s lineup of open-weight voice models, such as the compact yet powerful Parakeet, which has 1.1 billion parameters.
Nvidia’s latest voice AI release comes during a busy period for AI voice technology. Earlier this month, Mistral shipped Voxtral 3, ElevenLabs launched Realtime for instant speech in voice agents, and Salesforce released Launa-Omni, its voice model.