Objective evaluation of speech signal quality by the prediction of multiple foreground diagnostic acceptability measure attributes.

A methodology is described to objectively diagnose the quality of speech signals by predicting the perceptual detectability of a selected set of distortions. The distortions are a statistically selected subset of the broad number of distortions used in diagnostic acceptability measure (DAM) testing. The justification for such a methodology is established from the analysis of a set of speech signals representing a broad set of distortions and their respective DAM scores. At the heart of the ability to isolate and diagnose the perceptibility of the individual distortions is a physiologically motivated cochlear model. The philosophy and methodology is thus distinct from traditional objective measures that are typically designed to predict mean opinion scores (MOS) using well versed functional psychoacoustic models. Even so, a weighted sum of these objectively predicted set of distortions is able to predict accurate and robust MOS scores, even when the reference speech signals have been subject to the Lombard effect.