Convolutional Recurrent Neural Network With Attention Gates For Real-time Single-channel Speech Enhancement

In this paper, we incorporate the attention gates (AG) into the convolutional recurrent neural network (CRNN) to perform speech enhancement. The attention gates, which enhance important features and suppress irrelevant parts, can help the system effectively generate more accurate complex ratio mask (CRM). Because the model takes into account the phase information, better speech quality can be obtained. Since the parameters of the proposed model can be reduced to only 2.3M, the computational complexity is low, and the objective of real-time speech enhancement can be achieved.