It is important for designing a noise reduction algorithm to evaluate the quality of the noise-reduced speech signals accurately and efficiently. Subjective evaluation is ideal, but requires a time-consuming listening test with a lot of subjects. Then, objective distortion measures are often used for efficient evaluation. However, almost all the distortion measures do not consider the temporal variation of the speech distortion. In this paper, the temporal aspect of the short-term speech distortion is investigated based on higher-order statistics, that is, variance, skewness, and kurtosis. It is interesting that the skewness of the short-term SNR gives a good explanation for the discrepancy between subjective and objective evaluation.
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.