In order to solve the problems of poor separation and low signal quality of separated signals in two speakers separation experiments, we propose a deep clustering model based on bidirectional long short-term memory network (BLSTM), which adds phase information to speech signal processing and uses deep clustering to differentiate two speakers. The phase of the speech signal has a significant influence on the pitch performance that the naturalness of the separated speech has obviously improved after adding the phase information. Besides, we also improve the activation layer of the network by selecting a more effective ReLU activation function, which not only improves the separation effect but also accelerates the calculation speed.