Nvidia Unveils Nemotron 3.5 Transcribe: A New Open-Weight Transcription Model

Nvidia has unveiled Nemotron 3.5 Transcribe, a free, open-weight transcription model that supports up to 40 languages. The release also introduces version 2 of the Granary speech dataset, expanding Nvidia’s toolkit for audio understanding.

Key capabilities include translating between 31 languages, processing recordings up to one hour in length, and accurately identifying up to eight speakers in a single session. These features position the model as a strong alternative to existing ASR solutions, particularly on the Hugging Face Open ASR Leaderboard where it demonstrates competitive performance.

Designed for developers building applications for business meetings and medical transcription, Nemotron 3.5 Transcribe can be fine-tuned with Nvidia’s NeMo Speech framework to recognize industry-specific jargon in any language. The model’s architecture, with 30 billion parameters, is optimized for deployment across a wide range of hardware—from the Jetson AGX Thor robotics controller and DGX Spark chips powering next-generation laptops to Blackwell datacenter processors.

While the model excels at offline transcription, it currently does not support live streaming. Nevertheless, Nvidia positions this release as part of a broader strategy to strengthen its open-weight voice ecosystem, joining models like the compact Parakeet (1.1 billion parameters) and competing with recent entrants such as Mistral’s Voxtral 3, ElevenLabs’ Realtime, and Salesforce’s Launa-Omni.

Why it matters for GPU / AI infrastructure: The availability of a high-performance, open-weight transcription model like Nemotron 3.5 Transcribe lowers the barrier to entry for enterprises seeking to deploy AI-powered speech processing on cost-effective hardware. By running on a single chip—from edge devices to large-scale datacenters—Nvidia demonstrates how GPU-accelerated inference can be scaled across the entire hardware spectrum, enabling real-time transcription workflows that were previously limited to cloud-based solutions.