NVIDIA releases open-weight speech model for transcription and translation

NVIDIA released Nemotron 3.5 Transcribe, an open-weight speech model that combines transcription, speaker labeling, translation and answers to questions about recorded audio. The company says it transcribes 40 languages, including seven Indic languages, and accepts plain-language prompts with specialized vocabulary and formatting instructions. The model is available on Hugging Face in BF16, FP8 and NVFP4 formats, with support for RTX Spark and DGX Spark systems, Jetson AGX Thor devices and GeForce RTX 5090 GPUs. Developers can deploy NVFP4 versions through NeMo-Speech.cpp or vLLM. Open weights let developers choose whether to process audio on a desktop or in the cloud; NVIDIA's accuracy rankings are company-reported.