NVIDIA has released Nemotron 3.5 Transcribe, an open-weight model that can transcribe recorded speech, label speakers, translate audio and answer questions about it. The company has also released Granary v2, the multilingual data and curation pipeline used to prepare speech for the model's training.
Transcribe supports transcription in 40 languages, translation between English and 31 languages, and speaker attribution in 13 languages. It handles recordings up to an hour long and can label up to eight speakers. Developers can use natural-language instructions to request a transcript format or ask questions about a recording. Audio question answering is trained in English, and language support varies by task.
NVIDIA says Transcribe ranks first among open-weight models on the English track of the Hugging Face Open ASR Leaderboard, with a 5.33% average word error rate. In the model card's comparison, it recorded 4.62% mean word error rate across seven languages and 10.98% on long-form audio. On David AI's diarization benchmark, which measures speaker attribution across 21 languages, it had an 18.26% error rate, the lowest of 14 systems tested.
Granary v2 provides data for adaptation
Granary v2's public manifest covers 463,062 hours of speech across 48 languages and 16 source corpora. NVIDIA says those figures describe the dataset release, not Transcribe's full training mixture. The pipeline creates different transcript forms, task labels and quality metadata. Developers can use its public manifests or run the NeMo Curator pipeline on their own recordings or authorized public corpora. Source audio must be downloaded separately under each corpus's license.
NVIDIA also published a NeMo Speech fine-tuning workflow for adapting recognition to specialized vocabulary and transcript conventions. In a medical speech case study, fine-tuning reduced word error rate on 2,764 clips from held-out speakers from 22.44% with the best prompt to 12.17%, a 45.8% relative reduction. The study used a separate set of 855 clips to choose the prompt and checkpoints.
Nemotron 3.5 Transcribe and Granary v2 are available on Hugging Face and ModelScope. Developers can serve the current Transcribe checkpoint with vLLM. NVIDIA says quantized versions and local deployment with NeMo-Speech.cpp are coming soon. The model is built for recorded audio, rather than live streaming.
