MMSE estimator for speech enhancement considering the constructive and destructive interference of noise

Abstract

A new minimum mean square error (MMSE) estimator for discrete cosine transform (DCT) domain speech enhancement considering the two state possibilities of signal and noise DCT coefficients, the constructive and destructive interference, is proposed. Given proper identification of these events are possible, an optimum non-linear MMSE estimator can be derived considering the conditional events. Compared to the previously reported dual-gain multiplicative filters, the proposed estimator demonstrates superior theoretical and experimental performance with respect to mean square error (MSE) improvement and other objective quality measures.