Voice AI in vogue: Nvidia releases new open transcription model

Nvidia is pushing harder into voice AI. On Wednesday, it released Nemotron 3.5 Transcribe, a free open-weight model that can transcribe audio in up to 40 languages. It also put out version 2 of Granary, its open source speech dataset.

The pitch is broad, but the target is clear: developers building tools for business meetings and medical transcription. Nvidia says the model can translate 31 languages to and from English, handle recordings as long as an hour, and separate as many as eight speakers. It also claims the model beats rivals on benchmarks, including Hugging Face’s Open ASR Leaderboard.

There’s a practical angle here too. Nvidia says developers can combine Transcribe, the updated Granary dataset, and tools like its NeMo Speech model to tune apps for industry-specific jargon in each language. The catch is simple: this is for recorded audio, not live streams.

Nvidia also says the model, which has 30 billion parameters, can run on a single chip. That includes its Jetson AGX Thor robotics controller, DGX Spark chip for new laptops, and Blackwell data center chips. It joins other open-weight voice models from the company, including Parakeet, which Nvidia describes as both very small and powerful at 1.1 billion parameters.

The release lands in a crowded stretch for AI voice products. Mistral shipped Voxtral 3 earlier this month. ElevenLabs launched Realtime for instant speech in voice agents. Salesforce released Launa-Omni. Everyone wants a slice of speech, and Nvidia clearly intends to own a few more of the crumbs.