Nemotron 3.5 Transcribe: NVIDIA’s free model for recordings in 40 languages
The open model transcribes recordings of up to an hour, tells up to eight speakers apart and translates 31 languages. It doesn’t handle live audio.
On October 28, NVIDIA released Nemotron 3.5 Transcribe, a free, open-weight speech recognition model. Version 2 of its Granary speech dataset came out alongside it.
What’s new
- Languages: transcription in up to 40 languages.
- Translation: NVIDIA says it translates 31 languages to and from English.
- Recordings: up to one hour.
- Speakers: tells up to eight people apart.
- Size: 30B parameters, running on a single chip, from Jetson AGX Thor and DGX Spark to Blackwell datacenter chips.
NVIDIA claims the model beats rivals on benchmarks, including Hugging Face’s Open ASR Leaderboard, though TechCrunch doesn’t give the scores. Target uses include business meetings and medical transcription.
How to try it
The model is on Hugging Face. Using the updated Granary dataset and NeMo Speech tools, you can adapt it to industry jargon in each language.
Limitations
It only works with recordings, not live audio, so it won’t do for live captions. The coverage doesn’t name a specific license or list the languages, so check the model page for yours. Consider the hardware before planning a local deployment: this isn’t a lightweight laptop model.
Prepared with AI from public sources.