One Open Model for Multilingual Transcription, Speaker Attribution, Translation, and Audio QA with Top Leaderboard Results, Plus the Open Granary v2 Data Pipeline and Fine-Tuning Recipes to Adapt It
Even when a transcript captures every spoken word, users or downstream applications may still be left wondering who said what, how to translate it, or how to format domain-specific terminology. Multilingual calls, fast-paced meetings, and clinical dictations all require distinct outputs: translated text, speaker labels, or accurately spelled medical terms.
NVIDIA Nemotron 3.5 Transcribe is an open-weight, promptable model that turns speech recordings into highly usable text, now available on Hugging Face and ModelScope. It consolidates speech transcription, speaker attribution (diarization), translation, and audio QA into a single checkpoint: supporting transcription across 40 languages, bidirectional translation between English and 31 of those languages, and speaker attribution across 13 languages. It can process recordings up to an hour long and tag up to 8 speakers. Developers