NVIDIA Releases Nemotron 3.5 Transcribe Open-Weight Speech Model: Transcribes 40 Languages, With a 463,062-Hour Audio Dataset
NVIDIA released Nemotron 3.5 Transcribe on October 28, an open-weight speech model that combines multilingual transcription, speaker attribution, translation, and audio question answering in a single checkpoint. Alongside it, the company released Granary v2, an open speech dataset and curation pipeline whose public manifest covers 463,062 hours of audio across 48 languages.
According to NVIDIA's announcement, the model transcribes 40 languages, translates 31 of them to or from English, and attributes speakers in 13 languages; it handles recordings up to one hour long and labels up to eight speakers. The model uses a 34-billion-parameter hybrid mixture-of-experts decoder (30B-A4B) that activates roughly 4.3 billion parameters per token, and ships in BF16, FP8, and NVFP4 quantized weights.
NVIDIA says the model ranks first among open-weight models on the Hugging Face Open ASR English Leaderboard with a 5.33% average word error rate, and ties for first among open-weight models on Artificial Analysis Speech-to-Text at 2.8%.
The weights are released under the OpenMDW-1.1 license for commercial or non-commercial use, and are available via Hugging Face, GitHub through NVIDIA NeMo Speech, NGC, and build.nvidia.com. Hosted endpoints are offered by Baseten, DeepInfra, and Simplismart, with interactive demos on Hugging Face Spaces and ModelScope Studios.
The release is a speech-specific model, distinct from NVIDIA's earlier Nemotron multimodal agent models. Speech transcription is a common building block for super intelligence (SI) applications, including meeting notes, customer-service calls, and media captioning.
Source: GSM · USA (original Chinese edition)
Photo: 图：英伟达位于加利福尼亚州圣克拉拉的总部大楼。CC BY-SA 4.0，经 Wikimedia Commons。 / Photo: NVIDIA headquarters in Santa Clara, California. CC BY-SA 4.0, via Wikimedia Commons.
