Noise Estimation Based on Soft Decisions and Conditional Smoothing for Speech Enhancement

In this paper, the speech presence probability (SPP) based method for noise power spectral density (PSD) estimation is studied. It is shown that the SPP can be represented using a sigmoid function. This function offers more flexibility, since the slope and the mean of the sigmoid function can be adjusted independently for a better trade-off between noise overestimation and underestimation. Also, when the sigmoid function is mapped only to the a posteriori SNR with fixed priors, the decisions might be too soft for the noise PSD estimate to update the noise power. Thus, harder decisions based on conditional smoothing is employed on top of the sigmoid function. This results in a better performance in terms of noise tracking and speech quality when compared to the reference methods.