Skip to Main Content
Existing objective speech quality measurement algorithms still fall short of the measurement accuracy that can be obtained from subjective listening tests. We propose an approach that uses statistical data mining techniques to improve the accuracy of auditory-model based quality measurement algorithms. We present the design of a novel measurement algorithm using the multivariate adaptive regression splines (MARS) method. A large set of speech distortion features is first created. MARS is used to find a small set of features that provide the best estimate ("model") of speech quality. One appeal of the approach is that the model size can scale with the amount of speech data available for learning. In our simulations, the new algorithm furnishes significant performance improvement over PESQ (perceptual evaluation of speech quality).