Eight speakers. Forty languages. One hour of audio. Those are the headline numbers on Nvidia's newest speech tool, and the price is zero.
Nvidia released Nemotron 3.5 Transcribe on Wednesday. It's an open-weight model, free to download, that turns recorded speech into written text. TechCrunch noted that the company also shipped a second edition of Granary, its openly shared speech dataset. Anyone can grab the model and build on it, which is the whole point.
What Nemotron 3.5 Transcribe actually does
By Nvidia's account, the model can translate between English and 31 other languages in either direction. It can chew through a recording that runs a full hour, and it can label up to eight different voices in one file. The company also claims it beats competing models on public benchmarks, Hugging Face's Open ASR Leaderboard among them. Those numbers come from Nvidia itself, so I'd hold off until independent testers have had a go.
The obvious customers are developers building meeting tools and medical dictation software. Both are among the oldest paying uses of AI in offices and clinics, so Nvidia is aiming at a market that already exists instead of inventing one.
The speaker-separation feature deserves more credit than it usually gets. Transcribing words is one problem. Knowing who said them is another, and it's what turns a wall of text into minutes you can actually act on. Long-file support matters for the same reason: a board meeting or a deposition doesn't politely end after ten minutes.
Jargon is the quiet selling point. A cardiologist and a contracts lawyer don't talk alike, and a generic model stumbles on both. Nvidia's pitch is that developers can combine the model with the updated Granary data and the company's NeMo Speech toolkit, then tune an app to a particular industry's vocabulary, one language at a time. Think of it as sending the software to night school.
There's a real limit, though. This model reads recordings and doesn't handle live streams, so if you want captions appearing as someone talks, you'll need something else.
Small enough to run on one chip
Thirty billion parameters is not tiny, yet Nvidia says the model fits on a single chip. The list of compatible hardware is broad. At one end sit Blackwell servers. In the middle is the DGX Spark chip found in new laptops. At the edge is Jetson AGX Thor, a controller designed for robots. Picture a hospital that would rather not send patient audio off to someone else's cloud. Keeping everything in-house is the appeal, and open weights mean no licensing negotiation first.
The release joins Nvidia's other open voice models, including Parakeet, a compact 1.1-billion-parameter option. Nvidia, conveniently, also sells the hardware all of this runs on. Free software that makes people want more silicon is a tidy arrangement.
A crowded month for talking machines
Nvidia has company. Mistral launched Voxtral 3 earlier this month, ElevenLabs introduced Realtime for near-instant speech in voice agents, and Salesforce put out its own voice model, Launa-Omni. Four big launches in a few weeks suggest voice has graduated from demo material to a real contest.
Notice the split, though. ElevenLabs is chasing speed, with agents that answer without a pause. Nvidia's model goes for accuracy and breadth on files already recorded. Different bets on what people will pay for.
What to watch now: whether outside developers confirm the benchmark wins, and whether Nvidia follows up with a live-streaming version. Without one, the real-time crowd keeps a big opening.