Abstract: To improve the accuracy of loudspeaker abnormal sound classification and reduce the impact of the high spatial dimension and redundancy of the psychoacoustic model, a feature extraction method that combines the psychoacoustic model with the Maximal Relevance Minimal Redundancy (MRMR) algorithm was proposed. The response signals of four types of loudspeakers: normal, low level, circle touch, and air leakage were processed by the psychoacoustic model to conduct singular value decomposition, and the obtained singular values of the psychoacoustic energy were taken as the candidate feature set; then the MRMR algorithm was used to select the best feature from this feature set; and finally, the least squares support vector machine(LSSVM) model optimized by the whale optimization algorithm(WOA) was used for abnormal sound classification. Experimental results show that the recognition rate of this feature extraction method can reach 97.47%, which has better classification performance than the existing feature extraction algorithms.