On reliability of log-spectral distortion measure in speech quality estimation

Abstract

In this paper, the results of experimental studies of the reasons for the low reliability of the log-spectral distortion (LSD) measure for estimating the quality of speech signals limited in the frequency band are presented. It is shown that the LSD measure has an increased sensitivity not only to the speech formants, but also to the shape of frequency response of low-frequency filters limiting the frequency band. It is also shown that normalization by variance of band-limited signals leads to LSD measure anomalous behavior. Simple modification of the LSD estimate is proposed, which makes it possible to increase its statistical stability.