“…Thanks to the rapid development of deep learning in recent years, speech enhancement, separation and dereverberation have received remarkable progress and consistent improvements on public datasets, e.g., WSJ0-2mix [1], VoiceBank-DEMAND [2] and REVERB Challenge [3]. Various deep neural network (DNN) architectures are proposed and reported to achieve significant improvements on speech enhancement [4,5,6] or separation [7,8,9,10,11] tasks. However, most of the models mentioned above focus on individual enhancement or separation task and haven't considered the real-world environment that speech overlapping, directional/isotropic noise and reverberation may exist together, which leads us to consider adopting one universal model to cope with speech enhancement, separation and dereverberation simultaneously.…”