NVIDIA released an open-weight speech model for transcription, translation and speaker attribution, alongside Granary v2 data resources and fine-tuning recipes.
NVIDIA announced Nemotron 3.5 Transcribe on October 28, 2026, as an open-weight model for working with recorded speech. A single checkpoint supports transcription in 40 languages, translation to or from English in 31, speaker attribution in 13, and audio question answering. It can handle recordings up to an hour long and label as many as eight speakers. Developers can steer output with natural-language instructions, including requests for speaker labels or translation.
NVIDIA says the model is available through Hugging Face and ModelScope. The company reports that it ranks first among open-weight systems on several speech evaluations: its cited Hugging Face Open ASR results include 5.33% average word error rate (WER) for English, 4.62% across seven languages, and 10.98% on long-form audio. On Artificial Analysis’s speech-to-text measure, it ties for first among open-weight models at 2.8% AA-WER. These figures come from different tests and should not be treated as directly comparable.
The model combines a speech encoder with a hybrid Mamba–Transformer mixture-of-experts language model. Of its 30 billion total parameters, about 4 billion are active per token, according to NVIDIA. The company also reports a benchmark in which a forthcoming quantized version processed 100 hours of audio in under three minutes on one RTX PRO 6000 Blackwell Server Edition GPU; the test used 30-second clips at high concurrency. The available checkpoint can be served with vLLM, while local deployment through NeMo-Speech.cpp and quantized versions are described as forthcoming.
The release also includes Granary v2, a multilingual data collection and reusable curation pipeline, plus NeMo Speech fine-tuning resources. In a medical transcription case study, fine-tuning reduced WER on held-out speakers from 22.44% to 12.17%, a reported 45.8% relative reduction. On general English, however, WER rose slightly from 1.54% to 1.69%, underscoring that domain adaptation can involve trade-offs.