In this paper, we first start from extracting music features and analyze the extraction methods for time-domain, frequency-domain, and cepstrum-domain features of music. Next, the logistic regression model is used to recognize music and deal with two-class and multi-class classification problems. To extract music and text features, a CNN-SVM-based classification model is proposed using a deep learning teaching practice structure. Finally, feature selection experiments are carried out on the extracted raw feature set using various feature selection algorithms to optimize the feature set and reduce classifier computation. The results show that the average number of feature items selected using the improved correlation coefficient method is 64, and the number of feature items is mostly distributed between 58 and 72. The data verifies that the improved correlation coefficient method proposed can effectively extract the features of traditional music, improve the accuracy of traditional music classification, and thus promote the development of the fusion of traditional music culture and vocal music teaching in colleges and universities.