NVIDIA has announced the release of Nemotron 3.5 Transcribe, an open-weight speech model designed to convert recorded audio into text, offering capabilities in transcription, speaker attribution, translation, and audio question answering within a single checkpoint. This new model supports transcription in 40 languages, translation between 31 of those languages and English, and speaker attribution in 13 languages, handling recordings up to an hour in length and identifying up to eight speakers.
Alongside Nemotron 3.5 Transcribe, NVIDIA has also released Granary v2, an open dataset and data curation pipeline. This dataset comprises a public manifest of 463,062 hours of speech across 48 languages and 16 source corpora. The Granary v2 pipeline, built with NVIDIA NeMo Curator, prepares multilingual speech for model training, generating aligned transcript variants, task labels, and inspectable quality metadata. These figures represent the dataset release and not the complete training mixture for the Nemotron 3.5 Transcribe model.
Nemotron 3.5 Transcribe, with 30 billion parameters where approximately 4 billion are active per step, is engineered for efficiency, allowing it to run on a single GPU. The model has demonstrated strong performance on various benchmarks, ranking first among open models on the Hugging Face Open ASR Leaderboard for English, with an average word error rate of 5.33%. It also leads in speaker attribution benchmarks, achieving the lowest diarization error rate among 14 systems tested on a separate 21-language test.
NVIDIA has emphasized the promptable nature of Nemotron 3.5 Transcribe, allowing users to provide specific instructions to guide the model’s output. This feature enables developers to request different transcript formats, such as verbatim with or without filler words, or to adapt the model to specialized vocabulary through prompts or fine-tuning. In a medical transcription case study, fine-tuning Nemotron 3.5 Transcribe reduced the word error rate by 45.8% on held-out speakers.
The Granary v2 dataset and its accompanying reusable curation pipeline are designed to empower developers to adapt the model to their specific audio data. While Nemotron 3.5 Transcribe is built for recorded audio and does not support live streaming, its capabilities make it suitable for applications such as meeting notes, call archives, and medical dictation.
❤️ Support Independent Journalism
Your contribution keeps our reporting free, fearless, and accessible to everyone.
Or make a one-time donation
Secure via Razorpay • 12 monthly payments • Cancel anytime before next cycle
(We don't allow anyone to copy content. For Copyright or Use of Content related questions, visit here.)
The Chenab Times News Desk
