A perceptually motivated estimator for speech enhancement

Research in recent years has shown that many commonly used short time spectral amplitude estimators for speech enhancement (SE), such as Minimum Mean Square Error (MMSE) estimator, are not optimal since they do not pay attention to the perceptual effects of different components in the estimated speech. Based on this, some SE algorithms have been proposed using non-symmetric cost functions. In particular, an algorithm has been introduced based on a modification of Itakura-Saito (MIS) measure to preserve weak speech segments. Although this proposed method shows improvement in terms of preservation of low power segments of the speech, it provides less background noise attenuation compared to MMSE method. In this paper, we propose a modified algorithm and compare its performance with the MIS and MMSE methods. Experimental results show that the proposed method performs better than the other two methods as it attenuates the background noise and preserves the low power segments of the speech very well.