Voice AI in vogue: Nvidia releases new open transcription model
techcrunch ·
Nvidia has released Nemotron 3.5 Transcribe, a free and open-weight AI model designed for audio transcription in up to 40 languages. This release coincides with the launch of version 2 of Granary, an open-source speech dataset intended to help developers customize applications for specific industry jargon. The model boasts capabilities such as translating 31 languages to and from English, processing recordings up to one hour long, and distinguishing between eight different speakers. Nvidia claims that Nemotron 3.5 outperforms competitors on benchmarks like Hugging Face’s Open ASR Leaderboard. Although it currently supports only recorded audio rather than live streams, the 30-billion-parameter model is optimized to run efficiently on various hardware, from single chips to datacenter servers. This launch positions Nvidia within a competitive market that recently saw releases from Mistral, ElevenLabs, and Salesforce. The tools are primarily targeted at developers building solutions for business meetings and medical transcription.
techcrunch