Nvidia has released Nemotron 3.5 Transcribe, a free, open-weight speech-to-text model capable of transcribing audio in up to 40 languages and translating 31 languages to and from English. The company also launched version 2 of its open-source Granary speech dataset, targeting developers building business meeting and medical transcription applications.
Nvidia on Wednesday released Nemotron 3.5 Transcribe, a free, open-weight model that transcribes audio in up to 40 languages. It also released version 2 of Granary, its open-source speech dataset. Nvidia says Nemotron 3.5 can translate 31 languages to and from English, handle recordings up to an hour long, and distinguish up to eight speakers. The company claims the model beats rivals on benchmarks, including Hugging Face’s Open ASR Leaderboard.
The tools are designed for developers building apps for business meetings and medical transcription, two areas where AI has established a foothold in the workplace. Nvidia says developers can use the model, the updated Granary dataset, and tools like its NeMo Speech model to customise apps to recognise an industry’s jargon in each language. One limitation is that Transcribe works with recorded audio, not live streams.
The company says the model, at 30 billion parameters, can run on a single chip, from the Jetson AGX Thor robotics controller and DGX Spark chip powering new laptops to Blackwell datacenter chips. It joins Nvidia’s open-weight voice models, including the smaller Parakeet model with 1.1 billion parameters.
Nvidia’s latest voice entry comes during a busy period of AI voice releases: Mistral shipped Voxtral 3 earlier this month; ElevenLabs launched Realtime for instant speech in voice agents; and Salesforce released Launa-Omni, its voice model.
Analysis
Why This Matters
- Nvidia's entry into open transcription models could lower costs for developers building multilingual voice applications, especially in business and healthcare.
- Running a 30-billion-parameter model on a single chip, from edge devices to data centers, makes on-device and local transcription feasible, reducing reliance on cloud APIs.
- The release adds to a wave of voice AI products from competitors like Mistral, ElevenLabs, and Salesforce, intensifying the race for developer adoption.
Background
Nvidia, best known for its graphics processing units (GPUs), has been expanding into AI software and foundation models. Its Nemotron family of models covers language, vision, and speech. The company previously released Parakeet, a smaller voice model. The new Transcribe model targets the transcription market, which has grown alongside the rise of AI assistants and meeting summarisation tools.
What to Watch
- Developer adoption and benchmarks: Whether Nvidia’s model consistently outperforms rivals on standard speech recognition tests.
- Integration into third-party apps: Expect announcements from meeting note and medical transcription startups using the new tools.
- Nvidia’s next move: The company may release live-streaming capabilities or a smaller, faster variant for real-time use.
Checked
12 claims in this story checked against 1 source: 12 supported.
Show each claim
- SupportedNvidia released Nemotron 3.5 Transcribe as an open transcription model.AI News & Artificial Intelligence | TechCrunch: “Nvidia on Wednesday released Nemotron 3.5 Transcribe , a free, open-weight model”
- SupportedNemotron 3.5 Transcribe is free, open-weight and transcribes audio in up to 40 languages.AI News & Artificial Intelligence | TechCrunch: “a free, open-weight model that transcribes audio in up to 40 languages.”
- SupportedThe model can translate 31 languages to and from English.AI News & Artificial Intelligence | TechCrunch: “Nvidia boasts that Nemotron 3.5 can translate 31 languages to and from English”
- SupportedNvidia released version 2 of its open-source Granary speech dataset.AI News & Artificial Intelligence | TechCrunch: “It also released version 2 of Granary, its open-source speech dataset.”
- SupportedThe model can handle recordings up to an hour long.AI News & Artificial Intelligence | TechCrunch: “handle recordings up to an hour long”
- SupportedThe model can distinguish up to eight speakers.AI News & Artificial Intelligence | TechCrunch: “tell up to eight speakers apart.”
- SupportedNvidia claims the model beats rivals on benchmarks, including Hugging Face’s Open ASR Leaderboard.AI News & Artificial Intelligence | TechCrunch: “It claims to beat rivals on benchmarks, including Hugging Face’s Open ASR Leaderboard .”
- SupportedThe tools target developers building business-meeting and medical-transcription apps.AI News & Artificial Intelligence | TechCrunch: “These tools target developers building apps for business meetings and medical transcription”
- SupportedNvidia says developers can use the model, updated Granary dataset and NeMo Speech to customise apps to recognise industry jargon in each language.AI News & Artificial Intelligence | TechCrunch: “Nvidia says developers can use the model, the updated Granary dataset, and tools like its NeMo Speech model to customize apps to recognize an industry’s jargon in each language.”
- SupportedTranscribe works with recorded audio, not live streams.AI News & Artificial Intelligence | TechCrunch: “Transcribe works with recorded audio, not live streams.”
- SupportedNemotron 3.5 Transcribe has 30 billion parameters.AI News & Artificial Intelligence | TechCrunch: “Transcribe, at 30 billion parameters”
- SupportedNvidia says the model can run on a single chip.AI News & Artificial Intelligence | TechCrunch: “the company promises Transcribe, at 30 billion parameters, can run on a single chip”
