An Effective Audio Classification Method Based On Data Augmentation Strategy

Audio classification is one of the main research directions of computer hearing.It needs to classify different sound sources according to the inherent characteristics of sound.With the excellent performance of deep learning in various fields,some researchers try to use convolutional neural network to complete the task of audio classification.Compared with traditional methods,the accuracy of audio classification has improved,but there is still room for further research.To solve the problem that convolutional neural network has a high demand for data,we proposes a double audio data augmentation method,which can greatly increase the amount of audio data to meet the training of neural network model,so that the final classifier has a higher generalization ability.Aiming at the problem that the classification accuracy of convolutional neural network is unbalanced in classification,this thesis proposes a method to build hierarchical classification model group,which is to build multi-layer model,and train the error prone classification to optimize the final accuracy of audio classification.The specific operation depends on the confusion matrix of the classification model in the test data to get the error prone class.Based on the error prone class data,it retrains the original model to get the next level model,and optimizes the results of the upper level model through the lower level model,improving the accuracy of audio classification.In this thesis,DDA-IRRF algorithm is summarized based on the double audio data enhancement method,and the hierarchical classification model group method is applied to DDA-IRRF algorithm to form DDA-IRMG algorithm.Four open data sets such as esc-50 and a self collected data set from a laboratory are used to test the algorithm.The experimental results show that the dual audio data enhancement method and hierarchical classification model group method have a significant positive effect on improving the accuracy of audio classification,and the method has few limitations and good universality.