Nvidia released Nemotron 3.5 Transcribe, a free speech-to-text model. It handles 40 languages and translates 31 to and from English. It processes recordings up to one hour long. It can distinguish up to eight speakers. It targets meeting and medical transcription apps. It works only on recorded audio, not live streams. Nvidia also released Granary 2, an open speech dataset. The model runs on a single chip, from robotics controllers to datacenter hardware. It joins Nvidia’s existing open voice models. Nvidia claims it beats rivals on public benchmarks.
What changed
No free, open-weight model from Nvidia covered 40 languages for transcription.
What it unlocks
Transcribing recorded audio in 40 languages with speaker separation and industry jargon support.
- 40 languages
- 31 languages to and from English
- up to 1 hour recordings
- 8 speakers
What you need to act on it
- recorded audio, not livestreams
Sources