A New Speech Enhancement Approach Based on Progressive Deep Neural Networks

In this paper, a speech enhancement method based on the framework of progressive deep neural networks (PDNNs) is proposed to alleviate the recognition performance degradation of the automatic speech recognition (ASR) system in low signal-to-noise ratio (SNR) environments. It aims at decomposing a regression task into multiple subtasks, which are closely related to each other, to improve the system performance. Then the learning targets of these subtasks are designed with gradually increasing SNR gains. Furthermore, a post-processing module, which benefits from the rich information of the learning targets, is applied to further improve the system performance. Experimental results reveal that the proposed method can achieve improvements in both objective and subjective evaluations in low SNR environments when compared with the conventional deep neural network method.