A Progressive Enhancement Method for Noisy and Reverberant Speech

In this paper, a speech enhancement method based on the framework of progressive deep neural networks (PDNNs) is proposed for low signal-to-noise ratio (SNR) and highly reverberant environments. It aims at assisting the complicated regression task of mapping noisy and reverberant speech to clean speech by utilizing two independent tasks, which suppress reverberation and noises respectively. Furthermore, a progressive learning approach is used for each task, which brings intermediate learning targets to enhance system performances. Experimental results reveal that the proposed method can achieve improvements in both objective and subjective evaluations in low SNR and high reverberation time 60 (RT60) environments when compared with the conventional deep neural network-based method.