The key facts
- Nemotron 3.5 Transcribe is an open 30-billion-parameter model supporting up to 40 languages
- Translates 31 languages to and from English, handles recordings up to an hour, distinguishes up to 8 speakers
- Works only with recorded audio, not streaming; runs on a single chip — from Jetson AGX Thor to Blackwell
Nvidia releases open transcription model
Nvidia on Wednesday released Nemotron 3.5 Transcribe — a free open-weight model for transcribing audio in 40 languages, along with the second version of its open speech dataset Granary. According to the company, the model translates 31 languages to and from English, handles recordings up to an hour long, and distinguishes up to eight speakers. Nvidia also claims the model outperforms competitors on benchmarks, including the Open ASR Leaderboard on Hugging Face.
Who it's for and its limitations
The tools are aimed at developers of apps for transcribing business meetings and medical transcription — two established niches for AI in the workplace. Using the model, the updated Granary, and tools like NeMo Speech, developers can adapt applications to industry-specific terminology in each language. A significant limitation: Transcribe works with recorded audio, not live streams.
Hardware requirements and market context
The 30-billion-parameter model, according to Nvidia, can run on a single chip — from the Jetson AGX Thor robotics controller and the DGX Spark chip that powers new laptopsto data center Blackwell chips. It joins the company's lineup of open-weight voice models, which also includes the compact Parakeet (1.1 billion parameters).
A wave of voice AI releases
The release comes at a busy time for voice AI: Mistral earlier this month released Voxtral 3, ElevenLabs launched Realtime for instant speech in voice agents, and Salesforce unveiled its voice model Launa-Omni. Thus, Nemotron 3.5 Transcribe enters a market where several companies are simultaneously expanding their speech recognition and synthesis offerings.