Speech is one of the best natural and convenient intercommunication manners among humankind. Nowadays, speech processing technologies have been broadly used in many applied fields. In this paper, the main research focus is on the study of speech enhancement and separation, which is one of the key technologies when we try to put the speech processing into reality. Firstly, we introduce speech single as well as the neural network elementary theory and propose based on the BP neural network speech enhancement system modeling method. Secondly, we integrate voice feature extraction and summarize the speech cepstrum and noise cepstrum valuation for neural network training and learning in order to eliminate the noise. Finally, the experiment proved this method surpass traditional the speech enhancement algorithm. The simulation result shows that this speech enhancement system design method can save the running time and the effect is good.