Speech enhancement based on sequential noise estimation with a masking property

In this paper, we describe a method for enhancing speech that has been corrupted by additive background noise that varies in time. The proposed nonlinear spectral-subtraction approach is based on sequentially updating the estimated noise in each frame and then using an adaptive mask that is based on the properties of the human ear. The intent of this method is to use sequential estimation of the time-varying background noise in order to decrease the perceived noise in the speech segments. Furthermore, in order to enhance the speech elements obtained by using this masking effect, we use the noise-masking signal-to-noise ratio (NMSNR), which is calculated from the regression line, to adaptively control the scaling function. Results of experiments show that the proposed approach can efficiently remove additive noise due to various sources.