Acoustic scene classification method based on multi-stream convolution and data augmentation

An acoustic scene classification method based on multi-stream convolution and data augmentation was proposed for the problem of insufficient acoustic scene classification accuracy and generalization ability of existing models while applying single feature input. First, the specific working principles of the convolutional neural network and mixup data augmentation were introduced. Then, based on the network parallel input theory, a multi-stream convolutional neural network was designed to achieve multi-stream feature fusion, which consisted of feature extraction module and feature fusion module. Furthermore, to further promote the accuracy of model and reduce the probability of over-fitting problem, a multi-dimension mixup data generate method was applied to smooth the feature data. Finally, nine feature combination schemes were used to carry out acoustic scene classification experiments based on Urbansound8K, ESC50 and ESC10 datasets. Experimental results show that the accuracy of the model is 88.29%, 77.75 and 96.25% respectively, which verifies that the model has higher accuracy and stronger generalization ability when using this method for acoustic scene classification research. © 2022, Editorial Board of Journal of Huazhong University of Science and Technology. All right reserved.