Nvidia releases open Nemotron 3.5 Transcribe model
Nvidia has released Nemotron 3.5 Transcribe, a free open-weight model for transcribing audio in up to 40 languages, alongside version 2 of its open-source Granary speech dataset. The company says the 30-billion-parameter model can process recordings of up to one hour, identify as many as eight speakers and translate 31 languages to and from English.
The release targets developers building workplace applications for business meetings and medical transcription. Those uses are among the established enterprise settings for speech AI, where usable transcripts depend not only on speech recognition but also on handling multiple participants and specialised vocabulary.
Capabilities aimed at multilingual recorded audio
Nvidia says Nemotron 3.5 Transcribe outperforms competing systems on benchmarks, including Hugging Face’s Open ASR Leaderboard. Its stated language coverage spans transcription in up to 40 languages, while its translation capability covers 31 languages to and from English.
Speaker separation is another feature Nvidia is highlighting. The model is designed to distinguish up to eight speakers in a recording, which can be relevant to meeting transcripts and other multi-party audio. Its maximum recording length is one hour.
There is an important operational boundary: Nemotron 3.5 Transcribe works with recorded audio rather than live streams. Teams looking for immediate voice-agent responses or real-time transcription therefore need to account for that limitation when evaluating the model.
Customisation tools and deployment claims
Nvidia says developers can combine the model and Granary 2 with tools such as NeMo Speech to tailor applications to industry jargon in each supported language. This makes dataset and vocabulary adaptation part of the proposition rather than a separate speech-recognition task.
The company also says Nemotron 3.5 Transcribe can run on a single chip. Nvidia names the Jetson AGX Thor robotics controller, DGX Spark chip used in new laptops and Blackwell data-centre chips as deployment options. The model joins Nvidia’s existing open-weight voice portfolio, which includes Parakeet, a 1.1-billion-parameter model.
A crowded period for voice models
The announcement arrives during a period of frequent voice-AI launches. Mistral released Voxtral 3 earlier in the month, ElevenLabs introduced Realtime for instant speech in voice agents, and Salesforce released Launa-Omni. Funding for companies such as Gradium’s $100M voice AI funding round also reflects the contest to build and deploy voice-model platforms across business use cases.
For businesses, the immediate implication is to assess speech systems against the actual workflow: recorded versus live audio, required languages, speaker count, terminology adaptation and available hardware. Nvidia’s release adds an open-weight option for recorded-audio applications, but its stated live-streaming limitation remains a decisive selection criterion.