Nvidia's open-weight transcription model pressures paid speech APIs

Nvidia has released an open-weight speech model that developers can download and run themselves, which puts pressure on hosted transcription services. The company published Nemotron 3.5 Transcribe, a 30-billion-parameter model, on Hugging Face on 28 October. TechCrunch's Julie Bort describes the model as free and open-weight, but the article names no specific licence.

The model transcribes audio in up to 40 languages and translates 31 languages to and from English. It handles recordings up to one hour long and distinguishes up to eight speakers. It works on recorded audio only, not livestreams.

Deployment range is the commercial detail. Nvidia says the model runs on a single chip, from Jetson AGX Thor and DGX Spark up to Blackwell datacenter chips. A buyer can therefore run it on small edge hardware or in a datacenter without changing the model.

Nvidia says the model beats rivals on benchmarks, including Hugging Face's Open ASR Leaderboard. The article gives no scores, so the claim cannot yet be checked against numbers. Nvidia also released version 2 of Granary, its open-source speech dataset. Related to this launch, Nvidia's Parakeet is a multilingual open-weight voice model with 1.1 billion parameters.

The clearest beneficiaries are teams that want transcription on their own infrastructure. An open-weight model with a range of supported hardware gives them a credible alternative to paying per use for a hosted service. Nvidia also gains a foothold in speech, where its chips and its open models can reinforce each other.

Paid transcription vendors are the most exposed, because their pricing depends on customers not running the model themselves. Founders evaluating speech products should watch two things: the licence terms, which the article does not name, and the benchmark scores Nvidia has yet to publish.