Nvidia Drops a 30-Billion-Parameter Voice Model That Runs Anywhere

Nvidia wants to power every corporate meeting and medical chart with a single free model. The chip giant released Nemotron 3.5 Transcribe, a thirty-billion-parameter open-weight model that processes audio in forty languages. It handles hour-long recordings and separates up to eight speakers at once. Nvidia says the system outperforms rivals on benchmarks like Hugging Face's Open ASR Leaderboard.

Yet the system carries a notable constraint. Transcribe works exclusively with recorded audio rather than live streams, placing it behind competitors racing to build real-time voice agents. Even so, Nvidia engineered the model to run on a single chip, spanning hardware from the Jetson AGX Thor robotics controller to Blackwell datacenter hardware.

Our read is that Nvidia is playing a long infrastructure game. By giving away a heavy transcription engine alongside its updated Granary speech dataset, the company ensures developers build custom enterprise jargon tools on its proprietary silicon. Rivals like Mistral, ElevenLabs, and Salesforce are pushing rapid speech agents, but Nvidia is betting that owning the underlying transcription layer matters more than winning the immediate conversational race.