Speech enhancement with adaptive spectral estimators

Abstract

Common statistical estimators for speech enhancement rely on several assumptions about statistical properties of speech and noise processes. In real applications, these assumptions may not be always satisfied due to the effects of a nonstationary environment. In this work, we propose new robust spectral estimators for speech enhancement by incorporation of calculation of rank-order statistics to existing speech enhancement estimators. The proposed estimators are better adapted to nonstationary characteristics of speech signals and noise processes in real environments. By means of computer simulations, we show that the proposed estimators outperform the known estimators in terms of objective criteria of quality.