Nvidia debuts Nemotron 3.5 Transcribe, an open-weight model for multilingual audio transcription and translation, and Granary v2, its open-source speech dataset

Audio transcription and translation tools become significantly more accessible and adaptable for specialized needs.

Nvidia has released Nemotron 3.5 Transcribe, an open-weight AI model designed for multilingual audio transcription and translation. This model can process up to 100 hours of audio in 40 languages, label eight speakers, and answers questions about the content in a single pass. It is a 30-billion parameter mixture-of-experts model with 4 billion active parameters, allowing it to run on hardware ranging from edge devices like Jetson AGX Thor to data center GPUs such as Blackwell Server Edition. Nvidia also launched Granary v2, an expanded open-source dataset for speech AI, which contains over 463,000 hours of audio across 48 languages. The model's accuracy is high, with a baseline word error rate of 30.8%, reducible with prompting and fine-tuning, achieving 11.8% on specialized benchmarks. Developers can access and adapt Nemotron 3.5 Transcribe via a Hugging Face demo.

The release of Nemotron 3.5 Transcribe and Granary v2 signifies a move towards more integrated and adaptable AI solutions in the audio processing space. By offering an open-weight model capable of comprehensive audio analysis—transcription, speaker identification, and question answering—Nvidia is lowering the barrier to entry for sophisticated applications. The model's ability to run on diverse hardware, from edge to cloud, coupled with its high performance even with compression, suggests a broad potential for deployment across various industries, including customer service, media, and accessibility tools.

This development impacts developers and businesses by providing a powerful, customizable foundation for building advanced audio intelligence features. The open-source nature of both the model and the dataset encourages community contributions and fine-tuning for specific domains, potentially accelerating innovation in areas like real-time meeting analysis, multilingual content creation, and personalized voice assistants. The granular control offered through prompting and fine-tuning means users can tailor outputs for diverse use cases, from simple transcription to complex information extraction.