Au-M-ol: Medical Audio LLM Cuts Word Error Rate by 56%

Researchers have developed Au-M-ol, a multimodal architecture that extends Large Language Models with audio processing for clinical tasks. The model comprises an audio encoder for medical speech features, an adaptation layer mapping audio into LLM input space, and a pretrained LLM for transcription and language understanding. In experiments, Au-M-ol reduced Word Error Rate by 56% compared to state-of-the-art baselines on medical transcription, performing well in noisy environments, with domain-specific terminology, and across speaker variability. The work is published on arXiv (2604.23284).