NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt

NVIDIA’s new Nemotron 3.5 Transcribe is trying to be the kind of speech model that doesn’t stop at a clean transcript. It can transcribe, attribute speakers, translate, and answer questions about audio in one checkpoint, and it’s now on Hugging Face and ModelScope. The model covers 40 languages for transcription, translates 31 of them to or from English, and can label up to eight speakers in recordings as long as an hour. That makes it useful for meetings, classroom recordings, and messy audio where “what was said” is only the first problem.

The pitch is not just versatility. NVIDIA is leaning hard on leaderboard numbers. On Hugging Face’s Open ASR Leaderboard, Transcribe ranks first among open-weight models for English at 5.33% average word error rate. It also posts 4.62% mean WER across seven languages in the model card’s multilingual comparison, and 10.98% on long-form audio. On Artificial Analysis, it ties for first among open-weight models at 2.8% AA-WER while using about 4B active parameters, even though the full model has about 30B parameters. That active-parameter figure is the point: the model is large, but only part of it is turned on per token.

Speaker attribution is another place where NVIDIA wants the model to stand out. On David AI’s multilingual diarization benchmark, it ranks first among evaluated open-weight models and second overall on the combined transcription-plus-speaker task. On the diarization track itself, it has the lowest error rate across 21 languages, just ahead of NVIDIA’s own Nemotron 3 Diarization model. And on VoiceArena’s English full-session benchmark, it posts the lowest cpWER among 14 systems and drops fewer words than the rest. In other words: it is not just hearing speech, it is sorting out who said it, even when people talk over each other.

The model is built for recorded audio, not live streaming, but NVIDIA says it can run on a single GPU from edge gear to data-center systems. On one RTX PRO 6000 Blackwell Server Edition GPU with NVFP4 precision, it processes 100 hours of audio in under three minutes, measured on 30-second clips at high concurrency. Developers can run it with vLLM or locally with NeMo-Speech.cpp, and they can steer the output with natural-language prompts like “transcribe this recording and label each speaker” or “translate this audio into English.” The model card also includes task-by-language guidance, because support is not identical across tasks and audio question answering is trained in English.

The bigger story is Granary v2, the open data pipeline NVIDIA is pairing with the model. Granary v2 covers 463,062 hours across 48 languages and 16 source corpora in its public manifest, up from 25 languages in Granary v1. It produces several transcript forms, contextual ASR terms, translations, audio captions, and speech-grounded question-answer pairs, then filters out noisy or low-quality material. NVIDIA says this is the part that lets developers adapt the model to the wording, accents, and conventions they actually need instead of hoping a general transcript prompt will magically know the right spelling of every domain term.

That shows up most clearly in the medical case study. A prompt that specified English and the Latin alphabet cut WER from 30.8% to 23.8% on the full benchmark, and fine-tuning brought it down to 11.8%. On held-out speakers, the best-prompt baseline scored 22.44% WER and the adapted model reached 12.17%, with improvements also showing up in keyword WER and semantic WER. So the headline isn’t just that NVIDIA launched a stronger open speech model. It’s that the company is packaging the model, the data curation, and the adaptation recipe as one system.