Voice AI at Peak Popularity: Nvidia Releases New Open Transcription Model
Alongside the model, the second version of the open speech dataset Granary for tuning applications to industry-specific vocabulary has been introduced.
Nemotron 3.5 Transcribe works with recorded audio and converts it into text in 40 languages. The model supports recordings up to one hour in duration, translates 31 languages into English and vice versa, and can distinguish up to eight conversation participants. It does not process real-time streaming audio.
The tool is designed for developers of applications for business meetings and medical transcription. Using NeMo Speech, Granary, and the model itself, they will be able to customize systems for industry-specific vocabulary in different languages.
The company states that the model, with 30 billion parameters, can operate on a single chip—from the Jetson AGX Thor robot controller and the DGX Spark chip to Blackwell server solutions. It is also claimed to outperform competitors in tests, including the Open ASR Leaderboard.
