Nvidia has officially released Nemotron 3.5 Transcribe, a free and open-weight machine learning model capable of transcribing audio across an impressive array of up to 40 languages. Alongside this versatile new audio model, the tech giant also rolled out version 2 of Granary, its comprehensive open-source speech dataset. Together, these tools are engineered to meet the growing demands of modern developers who are building sophisticated voice-driven applications for critical enterprise environments, specifically targeting the bustling sectors of corporate business meetings and automated medical transcription.
The release underscores Nvidia’s ongoing commitment to expanding its open-weight AI offerings, giving developers robust resources to build specialized applications. According to the company’s technical specifications, Nemotron 3.5 Transcribe boasts advanced translation capabilities, allowing it to seamlessly translate 31 distinct languages both to and from English. Furthermore, the model is built to process extensive audio files, handling continuous recordings of up to one hour in length without losing context or accuracy.
Another standout feature of the new model is its advanced diarization capability, meaning it can accurately distinguish and tell apart up to eight different speakers in a single audio recording. Nvidia boasts that these capabilities allow Nemotron 3.5 to outperform several key rival models on prominent industry evaluations, including Hugging Face’s competitive Open ASR (Automatic Speech Recognition) Leaderboard.
Targeting Established Workplace Use Cases
Business meetings and medical transcription represent two of artificial intelligence’s most firmly established footholds in the modern workplace. By targeting these areas, Nvidia is providing developers with the foundational architecture needed to streamline administrative workflows, document patient interactions accurately, and extract actionable insights from corporate discussions.
To achieve maximum effectiveness in these professional settings, Nvidia explains that developers can leverage the Nemotron 3.5 Transcribe model alongside the newly updated Granary dataset and other proprietary tooling, such as the NeMo Speech model. This combination of resources enables teams to customize and fine-tune their applications so they can easily recognize highly specific industry jargon and specialized terminology across multiple languages.
Despite its advanced capabilities, the model does come with a notable operational limitation: Transcribe is currently designed to work exclusively with recorded audio files rather than live-streaming audio feeds. This design choice makes it ideal for post-processing workflows, such as generating meeting minutes or summarizing recorded doctor-patient consultations, while leaving real-time streaming applications to other specialized systems.
Flexible Deployment and High Efficiency
Weighing in at 30 billion parameters, Nemotron 3.5 Transcribe has been engineered for high efficiency, with Nvidia promising that the model is versatile enough to run efficiently on a single hardware chip. This scalability spans a wide range of Nvidia-powered hardware configurations, starting from edge devices like the Jetson AGX Thor robotics controller and the DGX Spark chip designed to power a new generation of AI-enabled laptops running revamped versions of Windows 11. At the high end of the infrastructure spectrum, the model is fully optimized to run smoothly on powerful Blackwell datacenter chips.
This robust newcomer joins Nvidia’s expanding family of open-weight voice models. It complements existing solutions in the company’s portfolio, including the remarkably small yet highly capable Parakeet model, which operates on 1.1 billion parameters and specializes in multilingual automatic speech recognition.
A Crowded Field in AI Voice Development
Nvidia’s introduction of Nemotron 3.5 Transcribe arrives during an exceptionally busy period for artificial intelligence voice technology releases, as major tech companies race to capture market share in spoken-language processing. The competitive landscape has shifted rapidly over recent weeks, marked by a flurry of high-profile announcements from rival AI developers and enterprise software giants.
Earlier this month, Mistral rolled out its Voxtral 3 model, adding to the growing selection of advanced audio tools available to the developer community. Simultaneously, ElevenLabs introduced Realtime, a new offering designed to deliver instant speech generation for interactive voice agents. Adding further momentum to the voice AI boom, Salesforce released Launa-Omni, its own proprietary voice model aimed at enterprise customer engagement.
As the market continues to evolve at a breakneck pace, Nvidia’s combination of the Nemotron 3.5 Transcribe model, the version 2 Granary speech dataset, and scalable hardware deployment options positions the company to play a central role in shaping the next generation of multilingual voice applications across industries.