A Method Of Environmental Sound Classification Based On Residual Networks And Data Augmentation

As sound plays a crucial role in our interaction with the environment,environmental sound classification has become an important research topic due to its application in the fields of road surveillance systems,intelligent housing system and emotional perception.In recent years,with the rise and development of deep learning,those methods based on deep learning have been widely used in environmental sound classification tasks,which lay a technical foundation for the study of environmental sound classification.However,the current convolutional neural network for environmental sound classification is difficult to extend the depth of the model.In addition,the relative scarcity of labeled data for environmental sound classification task is also an important reason for that convolutional neural network is difficult to be improved on a simpler model.Although some new datasets have been published in recent years,they are still much smaller than the datasets available for study,limiting the development of environmental sound classification technology.In order to solve these problems,in this paper,by studying the residual network and data augmentation,the optimization scheme is proposed and achieved high accuracy of environmental sound classification.The main works of this paper are as follows:1.An environmental sound classification method based on residual network and data augmentation is proposed.Firstly,time stretch and pitch shift are used to expand the sound data samples;Secondly,based on the characteristics of environment sound,the data samples are divided into frames,and Mel Frequency Cepstral Coefficient(MFCC)and their deltas are extracted as feature parameters;Finally,the extracted features are sent into the residual network model named Env Res Net constructed in this paper for classification.Env Res Net is compared with a variety of environmental sound classification algorithms.The experimental results show that the method based on residual network and data augmentation can achieve better classification effect.2.An environmental sound classification method based on multi-level residual network is proposed.Firstly,by analyzing the characteristics of the residual network and the multi-level residual network,a multi-level residual network named Mul-Env Res Net is built,and a better model strategy is selected to alleviate the gradient vanishing.The Mul-Env Res Net is optimized,and the momentum gradient descent method is used to accelerate the training process and save the training time.Then,the MFCC and their deltas of the environmental sound signals are used as inputs to the Mul-Env Res Net.A convolution layer is used to extract the local features,and then the multi-level residual block is used to extract the deeper information.The pooling layer is used to retain the main features,reduce the amount of parameters and calculation,and improve the generalization ability of the model.Finally,the features are fed into the full connection layer and then into the Softmax layer to classify the results.Compared with other convolutional neural network models used for environmental sound classification tasks,the proposed method greatly deepens the number of layers of network structure,thus it can extract deeper important features and improve the classification accuracy.Experimental results show that the proposed method is effective in classifying environmental sounds.