New to this? Read it in simple words
- Nvidia made a free AI model that turns speech into written text.
- It works in 40 languages and can tell up to eight speakers apart.
- Companies may use it in their own products and run it on their own computers.
- Nvidia says it is very accurate, but no one outside has tested it yet.
- Transcribe
- To write down what people say in a recording.
- Open-weight model
- An AI model whose files anyone can download and run themselves.
- Parameter
- One of the many numbers inside an AI model that it learned in training.
Nemotron is Nvidia’s family of open AI models. On October 28, Nvidia added a model for speech, together with a new version of its Granary speech data.
What the model does
Nemotron 3.5 Transcribe turns audio recordings into text in 40 languages. Nvidia says it translates 31 of them to or from English. It labels up to eight speakers, so a transcript shows who said what.
It works on recordings of up to one hour, not on live audio. Users can give it instructions in plain words, for example names or terms to spell correctly.
The model is large but fairly cheap to run. Its model card counts about 34 billion parameters, including the part that listens to audio. Only about 4 billion are active for each piece of text.
From Nvidia’s blog and model card of October 28. The numbers are Nvidia’s own.
Open, with a licence for business
The weights are free to download from Hugging Face under the OpenMDW-1.1 licence. The model card says it is “ready for commercial or non-commercial use.”
Granary v2 is a set of labels for 463,062 hours of speech in 48 languages, from 16 collections. It contains no audio. Developers fetch the recordings from each original collection, under that collection’s own licence.
The first version of Granary covered 25 European languages, TNW and SiliconANGLE note.
Nvidia’s claims
Nvidia says the model is the most accurate open-weight model on Hugging Face’s English speech leaderboard. Its blog and its model card give different error rates for that test, 5.33% and 4.71%.
Nvidia also says the model processed 100 hours of audio in under three minutes on one of its server chips. That test used a smaller, compressed version that is not out yet.
TechCrunch, SiliconANGLE and TNW reported the launch, but none of them tested the model.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1Primary source · October 28, 2026NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt Nvidia, on Hugging Face
- 2Primary source · October 28, 2026Model card: NVIDIA-Nemotron-3.5-Transcribe-30B-A4B Nvidia, on Hugging Face
- 3Research · October 28, 2026Voice AI in vogue: Nvidia releases new open transcription model TechCrunch
- 4
- 5Research · October 28, 2026Nvidia releases all-in-one adaptable Nemotron 3.5 audio transcription and translation model SiliconANGLE
The accuracy and speed figures are Nvidia’s own, and its blog and model card disagree on one of them. No outside test has checked the model yet.
Nvidia’s open Nemotron model transcribes 40 languages. Silicon AI News, October 29, 2026. https://siliconainews.com/stories/nvidia-nemotron-3-5-transcribe
Nvidia released a free open model that transcribes 40 languages for commercial use, though its accuracy claims are unchecked and inconsistent. How we rate the mood
The mood shows how the news frames the event for the tech and AI world: as an opportunity (6 and above) or as a loss (below 5). It is not our verdict on who is right.