Speech intelligibility prediction as a classification problem

Speech Intelligibility Prediction (SIP) algorithms are becoming increasingly popular for objective evaluation of speech processing algorithms and transmission systems. Most often, SIP algorithms aim to predict the average intelligibility of an average listener in some specific listening condition. In the present work, we instead consider the aim of predicting the intelligibility of single words. I.e. we attempt to predict whether or not a subject in a listening experiment was able to correctly repeat a particular word. We base the prediction on a noisy and potentially processed/degraded recording of the spoken word (as presented to a subject), as well as a clean reference recording of the spoken word. The problem can be treated as a supervised binary classification problem of predicting whether a specific word will or will not be understood. We investigate a number of different ways to extract features from the degraded and clean speech samples. The classification is carried out by means of Fisher discriminant analysis. Despite the large variability of speech intelligibility experiments, it is possible to obtain a considerable degree of predictive power.