Nvidia Releases Nemotron 3.5 Transcribe, an Open-Weight Speech Model for 40 Languages
The 30-billion-parameter model handles up to one hour of recorded audio and arrives with a second version of the Granary speech dataset.
Nvidia released Nemotron 3.5 Transcribe on Wednesday, a free, open-weight speech-to-text model that supports 40 languages, TechCrunch reported.
Key details
- Languages: transcription in 40 languages, plus translation between English and 31 other languages in either direction.
- Speakers: can tell apart up to eight speakers in a recording.
- Length: handles recordings up to one hour.
- Size: 30 billion parameters, small enough to run on a single chip, from Jetson AGX Thor and DGX Spark systems to Blackwell data center GPUs.
- Limitation: it works on recorded audio, not live streams.
- Distribution: the model is available on Hugging Face.
Nvidia claims the model outperforms rivals on benchmarks including Hugging Face's Open ASR Leaderboard, according to TechCrunch; those results were not independently confirmed.
Granary v2
Nvidia also released the second version of Granary, an open-source speech dataset. Developers can combine it with the company's NeMo Speech tools to adapt models to industry vocabulary in each language.
Context
Nvidia positions the model for uses such as business meeting and medical transcription apps. It joins Parakeet, a smaller multilingual model of 1.1 billion parameters, in the company's speech lineup. Nvidia's open-weight releases tie its software to its hardware, since the models are tuned to run across its product range.
TechCrunch's report did not specify the model's license terms; readers planning commercial use should check the license on the model's Hugging Face page.
