Nvidia releases Nemotron 3.5 Transcribe for 40 languages

Nvidia released Nemotron 3.5 Transcribe on Wednesday, TechCrunch reports. The model is free, open-weight, and transcribes audio in up to 40 languages. Nvidia says it can translate 31 of those languages to and from English.

The model handles audio up to one hour long. It can tell up to eight speakers apart. The company claims the model beats rivals on benchmarks, including Hugging Face's Open ASR Leaderboard, TechCrunch says.

What it fits and does not fit

The model is 30 billion parameters total, in a mixture-of-experts shape. On the Hugging Face model page, Nvidia says about 4.3 billion of those parameters are active at a time. TechCrunch reports the model runs on a single chip. The chip can be the Jetson AGX Thor robotics controller or the DGX Spark chip powering new laptops. It can also run on Blackwell datacenter chips.

The useful limit is in the TechCrunch piece: the model works on recorded audio, not live streams. So a meeting-notes product or a medical dictation product can use it. A live caption or a live customer-service agent still needs a different model. ElevenLabs is named by TechCrunch as a rival for real-time use.

The dataset released alongside it

Nvidia also released the second version of Granary, its open-source speech dataset, TechCrunch says. Nvidia's Hugging Face page for the new model is marked as last updated on 28 October. Developers can combine the model, the updated Granary dataset and Nvidia's NeMo Speech tools. Nvidia says the combination teaches the model each industry's own words.

Nemotron 3.5 Transcribe joins Nvidia's older open-weight voice models. One is Parakeet, at 1.1 billion parameters. TechCrunch names the two main business use cases: business meetings and medical transcription.

The crowded voice shelf

TechCrunch places the release inside a busy month for voice AI. Mistral released Voxtral 3. ElevenLabs launched Realtime for instant speech in voice agents. Salesforce released Launa-Omni, its voice model. Four open or near-open releases in a few weeks is more than any team can test by reading announcements.

A team building a voice feature this quarter now has a real decision to make. Pick a short clip of the audio your product actually sees, with its noise, its speakers and its vocabulary. Run it through Nemotron 3.5 Transcribe, Voxtral 3, Parakeet and whatever paid API you are already using. The winner is the one that gets your audio right at the price you can afford. Nvidia's claim about the leaderboard is Nvidia's claim. The one that matters is the one measured on your own audio.

Source

Voice AI in vogue: Nvidia releases new open transcription model, TechCrunch. The model is on Hugging Face.