Nvidia Gave Away a Speech AI Model Good Enough to Worry the Companies Charging for One
On October 28, Nvidia released Nemotron 3.5 Transcribe, a free, open-weight speech-to-text and translation model, alongside Granary v2, an updated open speech dataset. On paper it reads like routine model-card housekeeping: a mixture-of-experts model with roughly 30 billion total parameters and about 4 billion active per token, transcribing in 40 languages, translating 31 of them to and from English, distinguishing up to eight speakers in a recording, and handling clips up to an hour long. What makes it worth a second look is who it competes with, and why Nvidia is willing to give it away.
A model built to run everywhere Nvidia sells a chip
Nemotron 3.5 Transcribe is designed to run on a single Nvidia chip, and the range is the point: the same model scales from a Jetson AGX Thor robotics board up through the DGX Spark modules now showing up in laptops, to full Blackwell datacenter GPUs. Nvidia says it can churn through roughly 100 hours of audio in under three minutes on a Blackwell card. The model's language backbone comes from Nemotron 3.5 Lightning, an earlier Nvidia release from August, and it joins a lineage that already includes the Parakeet and Canary transcription models and April's Nemotron 3 Nano Omni. The pattern across all of them is the same: Nvidia isn't selling the software. It's seeding demand for the hardware that runs it.
Granary v2, the accompanying dataset, undercuts its own first version in an odd way: it holds 463,062 hours of speech across 48 languages, compared with roughly a million hours across 25 languages in the original Granary. Nvidia is trading raw volume for linguistic breadth, and the dataset doesn't bundle the audio itself — developers pull it from each original source under that source's own license, which keeps Nvidia's own distribution terms simpler than owning the audio outright.
The business this actually threatens
Transcription and translation are not a side feature for everyone. They're the entire product for a cluster of well-funded startups — ElevenLabs, Deepgram, AssemblyAI — that charge per minute or per API call for exactly what Nemotron 3.5 Transcribe now does for free, on hardware a developer may already own. Nvidia's own comparison point in its materials is xAI's Grok transcription service, priced around 10 cents an hour; a free, self-hostable alternative with a first-place ranking on Hugging Face's open ASR leaderboard is a direct shot at that pricing model, even if Nvidia never mentions a competitor by name.
The one gap is live audio. Nemotron 3.5 Transcribe works on recorded files, not streams, which leaves real-time voice-agent transcription — the use case ElevenLabs has been building toward with its Realtime product — outside what Nvidia just gave away. That's a meaningful carve-out: it means Nvidia is commoditizing the batch-transcription market (meeting notes, medical dictation, call-center archives) while leaving live, low-latency voice pipelines as a place where paid vendors can still differentiate, at least for now.
What the numbers actually show
Nvidia's own figures should be read as a vendor's benchmark, not an independent one. The company says the model ranks first among open models on Hugging Face's English ASR leaderboard with an average word error rate around 5.3 percent, and in a separate fine-tuning demonstration on medical speech from Indian doctors, error rates fell from roughly 31 percent to under 12 percent after about 14 hours of tuning on a single GPU. Those are two different tests measuring two different things — general-purpose English accuracy versus a narrow, accented, domain-specific fine-tune — and outside parties haven't yet reproduced either number. What's verifiable is the strategic logic: Nvidia is the one company in the AI stack that profits even when the model itself is free, because every download is also an advertisement for the silicon it was trained and tuned to run on.
That's the real story behind another open-weight model drop. Mistral, Liquid AI, and DeepSeek open-source models to compete for mindshare and enterprise deployments against closed labs. Nvidia open-sources models to compete against its own customers' alternative suppliers — the specialized AI startups whose entire business case depends on nobody giving away what they sell.