Nvidia Opens 30B Nemotron 3.5 Transcribe and 463K-Hour Granary v2 Speech Dataset Across 40 Languages

Nvidia released Nemotron 3.5 Transcribe, an open-weight 30B-parameter MoE model (4B active) that handles transcription, translation, speaker ID, and Q&A over the same audio across 40 languages. Nvidia says the model processes 100 hours of audio in under three minutes on a Blackwell GPU and reports a 30.8% WER baseline that drops to 11.8% when fine-tuned on medical speech. Alongside the model, Nvidia opened Granary v2, a 463,062-hour speech dataset covering 48 languages.