Assessing Segmental Impact for Objective Speech Quality Evaluation

Abstract

Accurately predicting speech quality is important for the design of new speech coding and processing algorithms to improve speech communication. Existing speech quality metrics are computed with all speech segments, and do not consider the contributions of various speech segments for quality evaluation. The present work utilized a speech-level based segmentation method to separate a speech signal into high-, middle- and low-level regions, and computed the quality measures only with selected speech segments. Subjective speech quality rating data from 120 noise-masked/suppressed conditions (processed by 14 single-channel noise-suppression algorithms) were correlated with the objective speech quality indices. Results showed that compared with the conventional implementation with all speech segments, using middle-level speech segments to compute speech quality index could yield an improved correlation coefficient in predicting subjective quality ratings for most quality measures, particularly for the measure of output signal-to-noise ratio. The findings of the present work may provide a new scheme to improve the performance of objective speech quality assessment based on the segmental contributions of speech signals.