On the Predictability of the Intelligibility of Speech to Hearing Impaired Listeners

What information do we need to know about listeners to predict their performance on a speech intelligibility task and how well can we predict intelligibility anyway? This paper performs a meta-analysis on two speech intelligibility studies of hearing-impaired listeners in which we evaluate different approaches to building a predictive model of intelligibility. The model has two components: a cochlear loss term based on a number of psychoacoustic measures of hearing, and a supracochlear loss term to explain residual performance variation. These models are trained using a method of cross-validation to determine how well they might perform on new listeners and new tasks. We found that cochlear loss could only explain 40% of the variability in performance across hearing-impaired listeners, while the supra-cochlear loss can account for a further 20-40% depending on the task. The combined cochlear and supra-cochlear loss terms allow good estimates of intelligibility scores in the data, with speech reception thresholds on a novel listening task being predictable to within 1dB on average.