Speech dereverberation using linear prediction with estimation of early speech spectral variance

Abstract

In this paper, we present a new dereverberation algorithm based on the weighted prediction error (WPE) method. In contrast to the conventional WPE method which alternatively estimates the reverberation prediction weights and early speech spectral variance, the proposed algorithm estimates the latter efficiently by employing a geometric spectral enhancement approach and a proper estimate for late reverberant spectral variance (LRSV). Hence, our algorithm does not require iterations to estimate the reverberation prediction weights nor needs alternation between the prediction weights and the spectral variance of early speech. Performance assessments demonstrate considerable improvements in terms of speech quality measures and computational load compared to previous WPE-based dereverberation methods.