Model-based estimation of late reverberant spectral variance using modified weighted prediction error method

In this paper, we propose a new approach to estimate the late reverberant spectral variance (LRSV) for speech dereverberation in the short-time Fourier transform (STFT) domain. Our approach uses a model-based scheme involving the estimation of a smoothing (shape) parameter and the reverberant-only component of speech. We propose to obtain the shape parameter by using estimates of the spectral variances of the direct-path and reverberant-only components of the speech, which in turn, can be calculated by smoothing coarse estimates of these two components. Furthermore, an accurate estimate of the reverberant-only component is obtained by means of a moving average scheme. In order to obtain the preliminary estimates of the direct-path and reverberant speech components, we employ a modified version of the weighted prediction error (WPE) method. In contrast to the original WPE method, the suggested modification is implemented for shorter processing blocks, each consisting of a number of STFT frames. This block-wise procedure allows for adaptation to moderate changes in environment and makes the proposed approach also suitable for time-varying acoustic scenarios. Performance evaluations with respect to previous LRSV estimation methods demonstrate the superiority of the proposed approach in both time-invariant and time-variant reverberant environments.