A framework for predicting speech quality using detectability of multiple distortions

摘要

This paper proposes a framework for predicting overall speech quality using a multi-dimensional analysis of individual distortions. The algorithm makes use of a physiologically motivated hydro-mechanical Cochlear Model to convert the speech signal into a domain that is more representative of what is perceived. Salient features are extracted and compared between the original and degraded representations to analyze the detectability of individual distortions. These are subsequently combined to predict overall quality.