Robust Speech Dereverberation Based on Adaptive Weighted Prediction Error Algorithm with Eigenvector Extraction

Abstract

Due to its satisfactory performance and no need for room impulse response information, the adaptive weighted prediction error (AWPE) algorithm is promising for speech dereverberation in practice. However, the robustness of AWPE to additive noise is low. To alleviate this problem, this paper proposes a variant of the AWPE algorithm that is based on eigen-decomposition of the signal auto-correlation matrix to construct the reference signal. By using the dominant eigenvector as the reference signal, a linear prediction filter is designed which has a better performance to predict the late reverberation even when the additive noise level is high. To reduce the computational complexity of the standard eigen-decomposition operation in the proposed AWPE variant, an online eigenvector extraction algorithm based on a fixed-point iteration algorithm is presented. Simulations are conducted to validate the effectiveness and robustness of the proposed algorithms over the standard AWPE algorithm.