Nvidia Launches Nemotron 3.5 Multilingual Audio Transcription Model

Nvidia released Nemotron 3.5 Transcribe, an open-weight 30-billion parameter mixture-of-experts model that handles speech detection, transcription, speaker labeling, and question answering in a single pass. It supports 40 languages, up to eight speakers, and can process 100 hours of audio. Out-of-the-box word error rate is 30.8%, reducible to 20.8% with prompting, and 11.8% when fine-tuned for specialized domains. The model runs on hardware ranging from edge devices to data-center GPUs. Nvidia also released Granary v2, an open-source multilingual speech dataset covering 463,062 hours across 48 languages.