Nvidia Expands Its AI Audio Suite With Nemotron 3.5 Transcribe and Granary Dataset 2

Nvidia has continued to cement its expansive footprint in the artificial intelligence landscape with the official release of Nemotron 3.5 Transcribe, a robust, free, and open-weight model designed to process and convert audio into text across an impressive portfolio of up to 40 languages. Alongside the rollout of this advanced transcription engine, the tech giant also introduced version 2 of Granary, its comprehensive, open-source speech dataset. Together, these tools represent a major strategic push by the company to capture a larger share of the enterprise and developer markets, specifically targeting high-value sectors where automated audio processing has become a foundational operational pillar.

The launch arrives at a time of intense competition within the artificial intelligence audio sector, following a flurry of high-profile voice model rollouts from rival companies throughout the month. Yet, Nvidia is positioning its latest offerings as uniquely capable, enterprise-ready solutions designed to address complex, multi-speaker environments and multi-language requirements with high precision.

Advanced Capabilities for Complex Audio Environments

According to technical specifications highlighted by Nvidia, the Nemotron 3.5 Transcribe model boasts a wide array of sophisticated features designed to handle demanding real-world audio scenarios. The system is capable of translating 31 distinct languages both to and from English, providing immense versatility for global organizations and international applications. Furthermore, the model can process exceptionally long audio recordings—spanning up to an hour in a single file—without sacrificing performance or accuracy.

One of the most notable technical highlights of the new model is its speaker diarization capability. Nemotron 3.5 Transcribe can accurately distinguish and attribute dialogue to up to eight distinct speakers in a single recording. This specific feature makes the model exceptionally well-suited for group settings where multiple individuals contribute to a conversation, such as corporate board meetings, legal depositions, panel discussions, and multi-party medical consultations.

In internal and comparative evaluations, Nvidia reports that Nemotron 3.5 Transcribe performs favorably against competing architectures, claiming top-tier performance on prominent industry evaluation benchmarks, including Hugging Face’s Open ASR Leaderboard. These benchmarks have long served as a crucial proving ground for automatic speech recognition systems, making Nvidia’s showing a strong statement regarding the model’s competitive edge.

Targeting Established Workplace Footholds

The release of Nemotron 3.5 Transcribe and the updated Granary dataset is strategically tailored for software developers and enterprise architects building specialized applications. The primary use cases for these tools center on business meetings and medical transcription, fields that represent two of the earliest and most firmly established footholds for artificial intelligence in the modern workplace.

Voice AI in vogue: Nvidia releases new open transcription model

In these environments, accuracy is paramount, and generic transcription models often struggle with industry-specific terminology, specialized jargon, acronyms, and regional accents. To overcome this hurdle, Nvidia emphasizes that developers can leverage the new model alongside the updated Granary dataset and other proprietary frameworks, such as the NeMo Speech model. This combination of resources allows engineering teams to fine-tune and customize their applications so they can effortlessly recognize and transcribe complex professional jargon across multiple languages.

Despite its robust capabilities, the model does come with certain operational boundaries. Industry observers and developers should note that Transcribe is currently engineered to work strictly with recorded audio files rather than live-streaming audio streams. This design choice aligns with its primary deployment targets, such as post-meeting summaries and archived medical dictations, where batch processing of recorded media is standard practice.

Hardware Versatility and Parameter Efficiency

Weighing in at 30 billion parameters, Nemotron 3.5 Transcribe strikes a notable balance between massive model capacity and operational efficiency. Despite its substantial size, Nvidia promises that the model is versatile enough to run efficiently on a single hardware chip. This scalability spans a broad spectrum of the company’s hardware ecosystem, ranging from edge computing devices like the Jetson AGX Thor robotics controller and the DGX Spark chip designed to power a new generation of AI laptops running revamped versions of Windows 11, all the way up to enterprise-grade Blackwell datacenter chips.

This architectural flexibility ensures that developers and enterprises can deploy the model wherever their infrastructure demands—whether locally on specialized edge hardware or at scale within massive cloud datacenters. Nemotron 3.5 Transcribe joins a growing family of open-weight voice models offered by Nvidia, sitting alongside smaller yet highly efficient offerings such as the 1.1-billion-parameter Parakeet model, which has also gained traction for multilingual automatic speech recognition tasks.

A Crowded Field of AI Voice Releases

Nvidia’s introduction of Nemotron 3.5 Transcribe takes place during a remarkably busy period for artificial intelligence voice and audio technology developments across the tech sector. The past several weeks have seen a rapid succession of competitive announcements from major players and specialized startups alike.

Earlier in the month, European AI leader Mistral released Voxtral 3, adding to its growing suite of multimodal capabilities. Concurrently, ElevenLabs announced the launch of its Realtime tool, designed to deliver instant speech generation for conversational voice agents. Salesforce also expanded its enterprise offerings with the introduction of Launa-Omni, its own proprietary voice model aimed at enterprise workflows.

This intense flurry of activity underscores the rapid commercialization and maturation of voice AI. As enterprises increasingly demand sophisticated, multilingual, and highly accurate audio tools to automate administrative and analytical workloads, Nvidia’s release of an open-weight 30-billion-parameter model and an updated open-source speech dataset provides developers with powerful new building blocks to compete in this fast-evolving market.