A Maximum Likelihood Approach to Deep Neural Network Based Speech Dereverberation

Recently, deep neural network (DNN) based speech dereverberation becomes popular with a standard minimum mean squared error (MMSE) criterion for learning the parameters. In this study, a probabilistic learning framework to estimate the DNN parameters for single-channel speech dereverberation is proposed. First, the statistical analysis shows that the prediction error vector at the DNN output well follows a unimodal density for each log-power spectral component. Accordingly, we present a maximum likelihood (ML) approach to DNN parameter learning by charactering the prediction error vector as a multivariate Gaussian density with a zero mean vector and an unknown co-variance matrix. Our experiments demonstrate that the proposed ML-based DNN learning can achieve a better generalization capability than MMSE-based DNN learning. And all the object measures of speech quality and intelligibility are consistently improved.