Nvidia has released Nemotron 3.5 Transcribe, an open speech model with 30 billion parameters that combines transcription, translation, and speaker recognition for up to 40 languages. The model processes recordings up to an hour long with up to eight speakers, responding to simple text instructions to handle tasks like translation, summarization, and speaker attribution.
Despite its 30 billion parameters, only about 4 billion are active per processing step, allowing the model to run efficiently on a single Nvidia GPU such as the RTX PRO 6000 Blackwell Server Edition. On benchmark tests, the model ranks near the top for accuracy and speaker attribution, outperforming several commercial speech services and securing first place on the English Open ASR Leaderboard among open-weight models as well as the VoiceArena speaker recognition test.
Alongside the model, Nvidia introduced the Granary v2 dataset containing 844,000 hours of audio across 48 languages, along with tools for developers to adapt the model to specialized fields like medicine. Testing on medical recordings demonstrated that targeted fine-tuning significantly reduced error rates, dropping from 30.8 to 11.8 percent. This release matters because it provides developers with an efficient, highly accurate, open-weight speech model capable of complex multimodal tasks on standard hardware.