Research Of Sound Classification Algorithm Based On Deep Learning

Sound is a ubiquitous natural phenomenon,which contains rich information and constantly enhances our understanding of the objective world.Because of the complexity and diversity of sound,sound classification has always been a challenging problem.With the development of deep learning technology,a large number of sound classification methods have emerged.These methods are mainly divided into two categories: one is to extract acoustic features from the original audio for the input of the classifier,and then the classifier classifies the audio.However,the classification performance of these methods largely depends on the effectiveness of the representative features extracted from the audio;the other is to model the original audio directly without feature extraction,and to learn sound features directly from the original audio sampling points through the neural network.However,the shortcomings of this kind of algorithm are also obvious.using the original audio sampling points as the input of the neural network,the neural network needs to go through a lot of calculation to get the effective features that can be used as the basis for sound classification,which leads to the slow operation of the algorithm.Therefore,in view of the shortcomings of the above two kinds of sound classification methods based on deep learning,the main research work of this paper is as follows:1)A sound classification algorithm based on Wavenet is proposed.Wavenet is a classic automatic speech generation network,which has the ability to model the original audio.In this paper,by improving the data input of Wavenet network,the arrangement of threshold convolution units and the internal structure of threshold convolution units,the amount of calculation of the network is greatly reduced,and it is applied to the field of sound classification.The algorithm achieves a classification accuracy of 90.2% on Urbansound8 k data sets.Experiments show that the algorithm is feasible and effective.2)A sound classification algorithm based on frame-level attention mechanism is proposed.Firstly,the log gamma spectrum feature of the original audio is extracted by convolution neural network,then the extracted features are convoluted,reduced in frequency domain and activated by sigmoid to get the time domain attention weight,then the log gamma spectrum feature is multiplied by the time domain attention weight to get the frame level attention feature of the original audio.Finally,the frame level attention feature is used as the input of the algorithm proposed in 1).As a result,a sound classification algorithm based on frame-level attention mechanism is obtained.The experimental results show that the accuracy of this algorithm is 1.9% higher than that of the algorithm in 1)on Urbansound8 k data sets.