Feature mining for GMM-based speech quality measurement

We propose a novel approach to objective speech quality measurement using feature mining and Gaussian mixture models (GMMs). A large pool of perceptual distortion features is extracted from the speech signal and data mining techniques are used to sift out the most relevant feature variables from the pool. We examine using multivariate adaptive regression splines (MARS), classification and regression trees (CART), a hybrid CART-MARS scheme, and the sequential forward selection (SFS) algorithm for data mining. For our speech databases, the SFS algorithm provides best performance with a five-feature, three-component GMM. A reduction of 21.7% in root-mean-squared mean opinion score estimation error is obtained in comparison with ITU-T P.862 PESQ.