The chipmaker released a technical case study demonstrating how adapting its latest speech-to-text model to specialized data can reduce mistakes in transcription. Domain-specific fine-tuning of Nemotron 3.5 Transcribe lowered the word error rate on medical recordings to 12.17% from a baseline of 22.44%, a performance gain achieved using about 14 GPU hours on a single GB300 processor. Nvidia is publishing the fine-tuning recipe alongside the model weights to provide developers a framework for tailoring the tool to specialized vocabularies.
Nemotron 3.5 Transcribe is a 30B parameter Mixture-of-Experts model with roughly 4B active parameters that can process audio in 40 languages and identify separate speakers in the same recording. Designed for non-streaming audio up to one hour, the open-weight tool currently ranks first among open-weight models on the Hugging Face Open ASR English leaderboard. It is available for commercial use under the OpenMDW-1.1 license.