Nvidia on Wednesday released Nemotron 3.5 Transcribe, a free open-weight model that transcribes audio in up to 40 languages. We also released version 2 of Granary, an open source audio dataset.
Nvidia boasts that the Nemotron 3.5 can translate 31 languages from English to English, handle up to an hour of recordings, and distinguish between up to eight speakers. It claims to outperform its rivals on benchmarks such as Hugging Face’s Open ASR Leaderboard.
These tools are targeted at developers building apps for business meetings and medical transcription, two of the most established footholds for AI in the workplace. Nvidia says developers can use tools like this model, the updated Granary dataset, and the NeMo Speech model to customize their apps to recognize jargon in each language. One drawback: Transcribe works with recorded audio, not live streams.
Still, the company promises that Transcribe can run in 30 billion parameters on a single chip, from the Jetson AGX Thor robot controllers and DGX Spark chips that power new laptops, to Blackwell data center chips. This joins Nvidia’s open-weight audio models, including the ultra-small but powerful Parakeet (1.1 billion parameters).
Nvidia’s latest voice entry comes during a busy period for AI voice releases. Mistral shipped Voxtral 3 earlier this month. Eleven Labs launches Realtime for instant speech on voice agents. Salesforce has released its voice model, Launa-Omni.
