NVIDIA Releases an Open Speech Model and the 463,062 Hours of Audio Behind It
On Oct. 28, 2026, NVIDIA released an open-weight speech model called Nemotron 3.5 Transcribe and published the audio dataset behind it as a free download. The company says the model transcribes 40 languages, translates 31 of them to or from English, and marks who is speaking in 13 of them. The dataset, Granary v2, covers 463,062 hours of speech across 48 languages.
Both the model weights and the dataset are posted on Hugging Face and ModelScope, so a developer can run transcription on their own hardware instead of sending audio to an outside service. NVIDIA describes the model as promptable: the same file produces a plain transcript, a translation or an answer to a question about a recording, depending on the instruction it is given in ordinary language.
According to the release post, the model handles recordings up to an hour long and labels up to eight speakers; it is built for recorded audio rather than live streaming. The published checkpoint holds 30 billion parameters but uses only about 4 billion of them per token. Quantized versions that would let it run locally on the company's desktop and embedded hardware are listed as coming soon.
On the Hugging Face Open ASR Leaderboard, which scores transcription accuracy on shared test sets, NVIDIA reports its model first among open-weight models on the English track with a 5.33% average word error rate, a comparison that its own chart notes excludes models sold only as paid services. Those placements, and the speaker-labeling results beside them, are NVIDIA's reading of outside leaderboards rather than a measurement made independently of the company.
Granary v2 ships as a set of manifests rather than as audio files, listing the transcript variants, labels and quality flags the pipeline produced; the recordings themselves have to be fetched from the original corpora under each one's own license. The release post also says language support differs by task, and that the model's question-answering mode was trained in English only.
