Nvidia releases a free model that transcribes audio in 40 languages
Nvidia tells that the model operates on one chip, but it works only with recorded audio and not live voice.
On Wednesday, Nvidia released Nemotron 3.5 Transcribe, a free model with open weights that you can download. It changes voice in a maximum of 40 languages into text and can identify a maximum of eight speakers. Nvidia also released Granary version 2, open voice data for apps about meetings and medical notes, two fields where voice AI is already common.
How it works
Other systems use different steps: find the voice, write it down, identify the speakers, then give information about it. Nvidia tells that Transcribe does all of this in one step. A user can tell it to translate the audio or to give the decisions made in a meeting.
The numbers
All numbers here are from Nvidia.
- The word error rate is 30.8% out of the box.
- With careful commands it falls to 20.8%.
- On a medical voice test from India, the adjusted model gets 11.8%.
- One Blackwell chip with compression does 100 hours in less than three minutes.
What is not known
Nvidia tells that its scores are above those of rivals on tests that are open to all. No rival scores are given. It works with recorded audio and not with live streams.
Insight: Nvidia gives away open models that operate on its own chips
On 21 October, Nvidia released open Nemotron 3.5 models that help agents find documents, and this release adds voice. The model operates on one Nvidia chip, and Cassava also added Nemotron models to its system. This reads as a move to make Nvidia chips the first selection for free AI parts. For builders: the model is free, but the chips are from Nvidia.
Sources
Posted
