Research And Application Of Music Classification Based On Convolutional Neural Network

The music information retrieval is an emerging field of the interdisciplinary research.In the past,the music relied on manual classification of music labels,but such classification was easy to retrieve errors because of mistakes made.In order to solve the problem of automatic classification,some people started to set out from the characteristics of the audio itself,and manually extract the audio features for automatic classification.However,since the classification of audio contains complex features and difficult to extract suitable audio features,the automatic classification of music is not particularly accurate.This paper devised a way to transform audio into spectrograms so that audio is handled in the form of pictures.At present,the convolution neural network is widely used in image recognition.In this paper,convolution neural network is also used to extract the deep features of the spectrogram,and music genre recognition is realized from the window size at different times.In order to better judge the style of music from the whole,this paper also set up a global pooling layer,which verifies the influence of different thresholds and different pooling sizes on the neural network.The experimental optimal parameters are found through experiments.In order to further improve the recognition accuracy of the system,this paper re-optimized the system architecture and proposed a circular convolution neural network.This architecture adds cyclic neural network to convolutional neural networks to process time-series data.In the training phase,through different parameter experiments,the optimal settings were found,and the dynamic weighted Euclidean loss function was designed for neural network training to further improve the stability of the system.