This paper proposes a neural network based speaker verification system. Different with the determined way of traditional method, the neural net support the ability of automatically constructing the separation space from the data we give. In this paper We just utilize the clustering power of neural net and the statistics of the speech's features onto the speaker verification system. A speech input is segmented into small frames and we takes their linear predicted coefficients (LPC) as their feature maps. Then a three layer perceptron is used as training and recalling machine. After training, the results of recalling test has shown that it can get very good performance (inside test 98.5% and outside test 98.3%)