Nvidia continues to aggressively expand its footprint in the artificial intelligence ecosystem, releasing Nemotron 3.5 Transcribe, a powerful, free, and open-weight model designed for advanced audio transcription across up to 40 languages. Alongside the model release, the hardware and AI giant also rolled out version 2 of Granary, its comprehensive open-source speech dataset. Together, these tools represent a major push by the company to capture a larger share of the enterprise productivity, corporate meeting, and medical documentation markets—areas where automated speech recognition has long served as a foundational technology.
The newly introduced Nemotron 3.5 Transcribe comes with a robust set of capabilities tailored for complex, multi-speaker environments. Nvidia boasts that the model can seamlessly translate 31 languages to and from English, process lengthy audio recordings spanning up to an hour, and accurately differentiate between as many as eight distinct speakers in a single audio file. Furthermore, the company claims that Nemotron 3.5 outperforms several competing models on prominent industry benchmarks, including Hugging Face’s highly competitive Open ASR Leaderboard.
By making the model open-weight and pairing it with updated training data, Nvidia is directly targeting software developers and enterprise architects who are building specialized applications for business meetings and medical transcription. These two verticals represent some of the most established and high-value footholds for artificial intelligence in the modern workplace. According to the company, developers can leverage Nemotron 3.5, the newly updated Granary dataset, and existing proprietary tools like the NeMo Speech model to heavily customize applications. This allows the software to recognize specific industry jargon, specialized terminology, and technical vocabulary in multiple languages. However, the model does come with a notable limitation: Transcribe is built strictly to handle recorded audio files rather than real-time, live-streamed audio feeds.
Despite its substantial scale—clocking in at 30 billion parameters—Nvidia emphasizes that Nemotron 3.5 Transcribe has been engineered to run efficiently on a single hardware chip. This versatility spans across the company’s hardware stack, ranging from the Jetson AGX Thor robotics controller and the DGX Spark chip designed to power a new generation of Windows 11 AI-enabled laptops, all the way up to heavy-duty Blackwell datacenter chips used by cloud providers and large enterprises. This new offering adds to Nvidia’s growing family of open-weight voice models, joining smaller yet capable alternatives like Parakeet, an efficient 1.1-billion-parameter multilingual automated speech recognition model available through Nvidia’s developer platforms.
Nvidia’s entry into the advanced voice transcription space arrives during an exceptionally busy period for generative AI audio releases across the tech sector. The competitive landscape has heated up significantly over recent weeks, with several major industry players rolling out their own voice and speech innovations. Earlier this month, Mistral introduced Voxtral 3, expanding its multimodal capabilities. Meanwhile, ElevenLabs launched its Realtime tool aimed at delivering instant speech generation for conversational voice agents, and enterprise software giant Salesforce debuted Launa-Omni, its own specialized voice model. As the market for voice AI accelerates, Nvidia’s combination of open-weight flexibility, high parameter scale, and broad hardware optimization positions the company as a formidable contender in the enterprise speech recognition arena.
