Evaluation of the parameters of spectral subtraction techniques and its effects on speech signal quality

Abstract

This article presents the results of tests conducted to evaluate various versions of the spectral subtraction technique. Brief descriptions of the versions of the techniques evaluated and the measures used to evaluate them are given. The tests were conducted using a set of signals, containing voices of male speakers, that were contaminated with white gaussian noise at different levels of contamination. A relation between the quality of the signals treated with some of the versions of the technique, and the value of the oversubtraction and spectral floor parameter, used on this versions, was observed in the results. The results show that an increase of the oversubtraction parameter, used to enhance the effect of the technique on the intelligibility of the speech signal, deteriorates the quality of the signals. Intellibility was evaluated using the Articulation Index while the quality was evaluated using the Itakura-Saito Distance and the Log-Likehood-Ratio.