The invention discloses an audio feature extraction method based on KL divergence. The method includes steps: reading audio files in a training set according to classes, converting the audio files tocorresponding power spectrums, obtaining an inter-class KL divergence matrix of different acoustic scenes in the training set through calculation, dividing frequency groups of a KL filter, and designing the KL filter; enabling the power spectrums of the training set to pass through the KL filter, and extracting frequency spectrum features based on the inter-class KL divergence; taking logarithms of the KL frequency spectrum features, performing normalization, and inputting the normalized features to a convolutional neural network for training to obtain an acoustic model; reading audio files ofa test set, converting the audio files to corresponding power spectrums, and extracting KL frequency spectrum features corresponding to the test set through the KL filter; and inputting the KL frequency spectrum features of the test set to the trained acoustic model for testing and assessment, and obtaining the final accuracy of the acoustic scene classification model. The method is applied to other acoustic scene classification events, and better performance compared with conventional Mel feature extraction method based on human ears can be obtained.