Speech Emotion Recognition (SER) has become a hot topic recently. In this paper, Back Propagation Neural Network (BPNN) was used as a training system for SER classification, and four emotional speeches of the German Berlin Emotional Database (EMO-DB) were selected as the experimental data-set. The recognition accuracy was compared under different number of nodes in the hidden layer, and the best classification model was determined by combining the training time and the mean squared error (MSE). The experimental results showed that when the number of nodes in the hidden layer is 14, the MSE is minimum, and the average recognition rate of BPNN reaches 98.81%. By compared with different number of nodes in the hidden layer, the average recognition rate increased by 0.2% to 23.5%.