Objective speech quality measures based on human audition

Abstract

To accurately assess the quality of a speech transmission or coding system generally requires subjective ratings obtained by experiment. An alternative measurement procedure which would be highly desirable would involve an objective assessment automatically performed by computer. Using a large data‐base consisting of many distortions as well as the corresponding quality ratings, we have attempted to predict subjective quality based on psychoacoustic models. Such properties as intensity perception, frequency resolution, and time integration were incorporated. One of the most interesting results involves the fact that quality of distortions involving waveform coding and additive noise were well modeled by psychoacoustic parameters. However, the quality of more complex distortions, such as those produced by linear predictive coding, was best predicted using parameters (e.g., growth of loudness) at substantial variance with psychoacoustic test results. One conclusion is that complex distortions are not well characterized by simple measures. However, many of our measures outperformed more traditional measures by a large margin. [Work supported by the Defense Communications Agency.]