Environmental sound classification based on parallel type neural network
Aiming at the problem of low accuracy of traditional single input model in environmental sound classification,a parallel feature fusion neural network based on time domain features and frequency domain features is proposed. In this network,firstly,the original audio is processed by data enhancement method;and then,the processed original audio data and Mel spectrum feature data are sent to the original waveform network and Mel spectrum network,respectively,after obtaining the time domain and spectrum domain features,the feature fusion is performed. Finally,the result is sent to SoftMax classifier for classification after feature fusion. Experimental verification is carried out on UrbanSound8K dataset,and the final classification accuracy is up to 96. 03％,which is prior to other models.
万方数据
维普
