Data driven method for non-intrusive speech intelligibility estimation

We propose a data driven, non-intrusive method for speech intelligibility estimation. We begin with a large set of speech signal specific features and use a dimensionality reduction approach based on correlation and principal component analysis to find the most relevant features for intelligibility prediction. These are then used to train a Gaussian mixture model from which the intelligibility of unseen data is inferred. Experimental results show that our method gives a correlation with subjective intelligibility of 0.92 and a correlation of 0.96 with the ANSI standard Speech Intelligibility Index.