Nvidia on Wednesday published Nemotron 3.5 Transcribe, an open-weight speech transcription model, and also released version 2 of its open-source speech dataset Granary. According to the company, Transcribe can transcribe recorded audio in up to 40 languages.
Capabilities described by Nvidia
Nvidia says Nemotron 3.5 can translate 31 languages to and from English, handle recordings up to one hour long, and distinguish as many as eight speakers on a single recording. The company claims the model outperforms competitors on several benchmarks, including rankings on the Hugging Face Open ASR Leaderboard.
One operational limitation: Transcribe works with recorded audio only and does not support livestream or real-time audio ingestion.
Technical details and deployment options
Nvidia lists Nemotron 3.5 as a 30 billion-parameter model and states it can run on a single chip. The company highlights support for hardware such as the Jetson AGX Thor robotics controller, the DGX Spark chip used in new laptops, and Blackwell datacenter GPUs.
The release complements Nvidia’s existing open-weight voice models, including the relatively small Parakeet model (1.1 billion parameters).
Target use cases and developer tooling
Nvidia positions the tools for developers building applications for business meetings and medical transcription, two established AI use cases in the workplace. The company says developers can use the model together with the updated Granary dataset and tools like its NeMo Speech model to adapt applications to recognize industry-specific jargon in multiple languages.
Market context
Nvidia’s announcement comes amid several recent AI voice releases: Mistral shipped Voxtral 3 earlier this month, ElevenLabs launched Realtime for instant speech in voice agents, and Salesforce released the Launa-Omni voice model.
