Abstract: Speech enhancement is a very important front-end work in speech signal processing and it directly affects the back-end's speech recognition effects. At present, the single-channel speech separation using neural networks has made great progress in solving cocktail party problem, but its separation effect on complex mixed speeches is not satisfactory. Aiming at the shortcoming of single channel, the multi-channel structure is adopted to form 4 super-directivity beams. The speech enhancement in a given direction is realized by using the multi-channel structure combined with neural network algorithm. The simulation and experimental results show that the proposed method has the obvious improvement in a variety of evaluation index compared with the super-directivity beamforming algorithm and spectral subtraction algorithm.