The topic of this paper is audio description (AD) for blind and partially sighted people. The author outlines a discourse-based approach to AD focussing on the role of mental modelling, local and global coherence, and different types of inferences (explicatures and implicatures). Applying these concepts to AD, the author discusses initial insights and outlines questions for empirical research. The author's main aim is to show that a discourse-based approach to AD can provide an informed framework for research, training and practice.