Free, open-weight speech models are getting good enough that paid transcription APIs should be nervous. Nvidia just added to that pressure. The company released Nemotron 3.5 Transcribe on Wednesday, a 30 billion parameter model that transcribes audio in up to 40 languages, along with version 2 of its open-source Granary speech dataset.
The feature list is solid. Nvidia says the model translates 31 languages to and from English, handles recordings up to an hour long, and separates up to eight speakers. It also claims to beat rivals on benchmarks, including Hugging Face’s Open ASR Leaderboard. Treat that last claim carefully. Vendors pick the benchmarks they win, and real-world accuracy on noisy meeting audio or accented medical dictation is what developers will actually test.
The target use cases are meetings and medical transcription, two of the most established places AI already works in offices and clinics. Nvidia pitches customization as the hook. Developers can combine the model, Granary, and tools like NeMo to teach apps an industry’s jargon in each language. That matters because generic transcription tends to fail on drug names, legal terms, and product acronyms. Owning the weights lets a team fine-tune for that, which a closed API usually doesn’t allow.
There is one clear drawback. Transcribe works on recorded audio, not live streams. For post-call notes and meeting summaries, that’s fine. For voice agents or live captioning, it’s a dealbreaker, and developers will look elsewhere.
Still, the deployment story is strong. Nvidia says the model runs on a single chip, from the Jetson AGX Thor robotics controller and the DGX Spark chip in new laptops up to Blackwell datacenter hardware. That’s a useful range for teams that need on-device or on-premises processing for privacy reasons, which is common in healthcare. It also sits alongside Parakeet, Nvidia’s much smaller open voice model at 1.1 billion parameters, giving developers a choice between accuracy and footprint.
The timing says more than the model does. Mistral shipped Voxtral 3 earlier this month. ElevenLabs launched Realtime for instant speech in voice agents. Salesforce released its own voice model, Launa-Omni. Voice is crowded, and the competition is split: ElevenLabs is going after live interaction, while Nvidia is filling the batch-transcription slot with free weights.
Nvidia’s motive is simple. Every open model that runs well on its chips sells more chips. So does this change anything? For startups paying per-minute transcription fees, yes, it’s worth a benchmark run on your own audio. For anyone building real-time voice products, it’s mostly noise for now.
