Relationship between subjective and objective evaluation of noise-reduced speech with various widths of temporal windows

Abstract

It is necessary to enhance adverse speech signals for building useful speech interfaces. Speech enhancement is essential under noisy and reverberant acoustic environments. Therefore, quality assessment of the enhanced speech signals should be also an important issue in noise reduction and dereverberation. Subjective evaluation is given by carrying out listening tests, and objective evaluation is provided by speech distortion measures. However, there is the discrepancy between subjective and objective evaluation of speech distortion. The author has investigated the relationship between subjective and objective evaluation of noise-reduced speech signals. The objective speech distortion was calculated in each short-term frame, of which length was fixed, and the statistical characteristics of the short-term speech distortion were investigated using higher-order statistics such as skewness and kurtosis. The preliminary result suggested that skewness of the short-term speech distortion could give an explanation for the discrepancy between subjective and objective evaluation. Further investigation of the relationship between subjective and objective evaluation of noise-reduced speech signals is carried out with a variety of temporal window widths. [Work supported by NEDO, Japan.]