Classification of Speech Degradations at Network Endpoints Using Psychoacoustic Features

We propose a method of classifying speech degradations at network endpoints. Perceptual features are extracted from degraded speech signals and used to form statistical reference models of behaviors of different degradation types. Consistency values between degraded speech signals and the reference models are calculated and used to train a degradation classifier. The consistency values of a received degraded speech signal then serve as predictors in the trained classifier for a degradation type decision. The proposed method is tested on four commonly encountered degradation types with unseen data and the experimental results show that the method achieves high classification accuracy. The proposed method can be used to enhance applications such as speech enhancement, recognition, and quality estimation.