A deep convolutional encoder-decoder model for robust speech dereverberation
Dawei Wang, Y. X. Zou, Weimin Shi · 2017
Research shows that speech dereverberation (SD) with Deep Neural Network (DNN) achieves the state-of-the-art results by learning spectral mapping, which, simultaneously, lacks the characterization of the local temporal spectral structures (LTSS) of speech signal and calls for a large storage space that is impractical in real applications. Contrarily, the Convolutional Neural Network (CNN) offers a better modeling ability by considering local patterns and has less parameters with its weights sharing property, which motivates us to employ the CNN for SD task. In this paper, to our knowledge, a Deep Convolutional Encoder-Decoder (DCED) model is proposed for the first time in dealing with the SD task (DCED-SD), where the advantage of the DCED-SD model lies in its powerful LTSS modeling capability via convolutional encoder-decoder layers with smaller storage requirement. By taking the reverberant and anechoic spectrum as training pairs, the proposed DCED-SD is well-trained in a supervised manner with less convergence time. Additionally, the DCED-SD model size is 23 times smaller than the size of DNN-SD model with better performance achieved. By using the simulated and real-recorded data, extensive experiments have been conducted to demonstrate the superiority of DCED-based SD method over the DNN-based SD method under different unseen reverberant conditions.