Objective measures of speech quality (subjective)

This thesis investigates objective measures of speech quality, or measures which can be computed from properties of an original and a distorted speech waveform. Two sets of objective measures are investigated. The first set is designed to estimate the specific types of perceived distortions specified by the 'parametric' subjective quality scales of the Diagnostic Acceptability Measure (DAM), a subjective speech quality test. The narrow scope of these parametric scales promotes a close coupling between physical quantities and their associated perceptual qualities, resulting in quite accurate measures. The second set of objective measures estimate the composite acceptability scale of the DAM, a scale measuring overall speech acceptability. The first set of measures are used as a foundation for designing the second set of measures. The speech quality data base used in this research is quite extensive. It consists of a total of 1056 examples of distorted speech produced by various speech coder and other speech distorting systems and their associated subjective quality ratings as produced by the Diagnostic Acceptability Measure. The thesis research can be divided into three parts. In the first part, the relationship between the scales of the Diagnostic Acceptability Measure and several of the best available objective speech quality measures is analyzed. It is demonstrated, using Multidimensional Scaling, that current objective measures produce poor estimates of composite acceptability because they incorporate little of the information provided by the parametric subjective measures. Multiple linear regression is used to find a linear relationship between the parametric subjective quality scales and the composite acceptability scale. The resulting regression model produces very good estimates of composite acceptability using only a subset of the parametric qualities. Because of this relationship, objective measures of parametric quality can be used as the basis for a measure of composite acceptability. In the second part, a set of objective measures is designed which provide dramatically improved estimates of the parametric qualities of the Diagnostic Acceptability Measure. Performance is measured in terms of the correlation between actual and estimated subjective quality. Multidimensional scaling graphically shows the remarkable ability of these measures to estimate the parametric subjective qualities. And finally, in the third part, the parametric objective measures are combined into a single composite measure for estimating subjective composite acceptability. Several measures are presented, with correlations to composite acceptability ranging from 0.81 to 0.85.