NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt
huggingface ·
NVIDIA has released Nemotron 3.5 Transcribe, an open-weight model that combines multilingual transcription, speaker attribution, translation, and audio question answering into a single checkpoint. The model supports 40 languages for transcription and achieves top rankings on several major leaderboards, including the Hugging Face Open ASR Leaderboard and David AI’s diarization benchmark. A key feature is its promptability, allowing developers to guide output formats and domain-specific jargon using natural language instructions without immediate fine-tuning. To support adaptation, NVIDIA also introduced Granary v2, an open data curation pipeline with over 463,000 hours of multilingual speech data. Case studies demonstrate that fine-tuning with this pipeline can significantly reduce word error rates in specialized fields like healthcare. The system runs efficiently on single NVIDIA GPUs, ranging from edge devices to data-center servers, utilizing a hybrid Mamba-Transformer architecture. Developers can access the model via Hugging Face, ModelScope, or hosted endpoints, with tools provided for local deployment and custom training.
huggingface