ICCRN: Inplace Cepstral Convolutional Recurrent Neural Network for Monaural Speech Enhancement

Abstract

According to the mechanism of speech production, speech can be decomposed into excitation and vocal tract which are sparsely represented in cepstral domain. In this study, we propose a neural network for monaural speech enhancement on time-frequency cepstral space that is implemented by inserting a cepstral frequency block into our inplace convolutional recurrent network. The proposed method has a good ability of restoring the speech masked by noise. Experimental results show that the proposed ICCRN model significantly outperforms the baseline system, particularly under low SNR conditions.