NVIDIA Releases Open-Source Speech Models Nemotron 3.5 Transcribe and Granary v2

On October 28, 2026, Hugging Face's blog (first-hand report) announced that NVIDIA released an open-weight speech model, Nemotron 3.5 Transcribe. The model integrates speech transcription, speaker diarization, translation, and audio question answering into a single model, supporting transcription in 40 languages, translation between 31 languages and English, speaker labeling in 13 languages, and can process recordings up to 1 hour long with up to 8 speakers, with tasks specifiable via prompts. The same report also mentioned that NVIDIA released the Granary v2 data pipeline alongside it. That evening, TechCrunch and The Decoder followed up with reports consistent with the first-hand report: TechCrunch described the model as freely available with open weights, supporting transcription in up to 40 languages, and released alongside the open speech dataset Granary 2; The Decoder described the same model as capable of transcription, English translation, speaker identification, and audio Q&A via a single prompt. None of the reports contradict one another. The event is still in its announcement stage, with no further follow-up reports yet.

AI overview of reporting · Updated 2 h前

Not enough continuous observations to show a trend yet.