Nvidia launches open-weight AI model for transcription in 40 languages
Nemotron 3.5 Transcribe also supports translation for 31 languages. Nvidia paired the launch with a speech dataset covering 48 languages.
Nvidia has launched Nemotron 3.5 Transcribe, a free, open-weight speech model that the company says can produce transcripts in 40 languages. It also released Granary v2, a free dataset containing 463,062 hours of speech spanning 48 languages.
The model supports translation between English and 31 languages, speaker identification and questions about recorded audio. Nvidia says it can distinguish as many as eight speakers and process recordings lasting up to an hour, but is not designed for live transcription.
Developers can use natural-language instructions to choose a verbatim or cleaned-up transcript and specify medical vocabulary or product terminology. Commercial use is permitted under the model’s license.
Nvidia says the model runs on one Nvidia chip, including Jetson boards and data-center GPUs. The company also claims the top position among open models on Hugging Face’s English speech-recognition leaderboard, reporting a 5.33% average word error rate.
In a customization demonstration, Nvidia trained the model using medical speech from Indian doctors. The company reported a 45.8% reduction in errors on held-out test clips after about 14 hours of training on one GPU.
Granary v2 expands language coverage from 25 in its predecessor to 48. Developers obtain the audio from its original sources, with each source’s license governing its use.
