DNN-Based Linear Prediction Residual Enhancement for Speech Dereverberation

In daily-life scenarios, reverberation inevitably causes a decrease in speech recognizability and speech quality. Exploring methods to eliminate reverberation will benefit both human perception and other speech technology applications such as identity authentication and speech recognition. This paper proposes a speech dereverberation algorithm based on linear prediction (LP) residual processing using deep neural network (DNN). The amplitude spectrum of the LP residual of short-term speech is used as a speech feature to train the DNN, and the mapping relationship between LP residual of the reverberant speech and that of the clean speech is learned. Comparative experiments under different reverberation conditions have verified the effectiveness and robustness of the algorithm.