| The audio scene classification task is to process the recorded audio data,extract the features in the audio data,and use the classifier to match the audio data with several pre-defined audio scenes after the processing of deep learning models.The application scenarios of audio scene classification are very rich and the demand for deployment on smart mobile devices is increasing,while the storage and computational resources available in mobile devices are limited,which poses a challenge for the deployment of audio scene classification systems in mobile.Low-complexity audio scene classification systems can be deployed in mobile smart devices,and in this paper,in order to investigate the effect of the ratio of the number of active channels to the total number of channels in the audio scene classification system on the model efficiency.A structured pruning method using pruning model channels based on scaling factor control is proposed to trim the redundant channels in the network.It is found that in the deep learning-based audio scene classification system,the pruning method is used to reduce the number of parameters in the model by increasing the channel activity in the model,and without affecting the model accuracy,the model efficiency is highest with high channel activity in the model and little loss of accuracy in the model when the pruning threshold is set to 50%.In this paper,three different network architectures are analyzed,namely VGG network,ResNet network and CRNN network.The proposed method can effectively compress the model on all three different architectures,and the model is most efficient when the channel activity is 60%.In order to study the effect of the ratio of the number of active channels to the total number of channels in the audio scene classification system on the efficiency of the model,this paper uses the pruning method to reduce the redundant channels in the model,which is used to maintain the performance of the model by increasing the channel activity in the model.In this paper,a low-complexity audio scene classification system that can be easily deployed in mobile smart devices is implemented by using a scaling factor-based pruning method,which reduces the number of model parameters with little loss of model accuracy. |
