Research On Music Genre Classification Model Based On Convolutional Neural Network

Music genre classification is a key link in music information retrieval.Different users have different preferences for different genres of music.No matter what genre,the composition of a piece of music is very complex,accompanied by a variety of instruments,vocal differences are also very significant,the harmony of various elements is ever-changing.Constructing a good music classification system can effectively reduce user's time-consuming for music retrieval and improve user experience.Early music genres were classified mainly through professional audio annotations,which undoubtedly took time and effort.After introducing the machine learning method,the possible acoustic features are determined by manual judgment,and the classifiers are trained by extracting these features in music,thus realizing the music genre classification.This method is unstable and needs to design feature sets manually,so it relies on personal experience judgement and professional knowledge to a certain extent,so it is difficult to improve the accuracy.To solve the above problems,based on the idea of deep learning and the structural characteristics of convolutional neural network,this paper designs a music genre classification model with spectrum as input,and provides a new idea of audio classification and recognition.The main tasks are as follows:1.The corresponding spectrograms of music files are generated by short-time Fourier transform,Meier transform and constant Q transform.The visual acoustical characteristics of these three kinds of music files are studied,and the visual differences of the image acoustical characteristics among them are studied.The maps generated by different schools are compared horizontally.A complete structure is designed from input to output,and a classification model of convolution neural network based on spectrum graph is built.The efficiency and powerful feature learning and classification ability of convolution neural network are utilized to reduce the time and cost of manual processing.On GTZAN(George Tzanetakis)dataset,the model has 71.34% classification accuracy,which is ahead of other common machine learning algorithms.2.To overcome the shortcomings of the original model in terms of input data and network architecture,an improved convolutional neural network classification model is proposed in this paper.The classification accuracy of the new model on GTZAN dataset is 92.21%,which improves significantly compared with the original model.In the future,the performance of the model will be further optimized in the aspects of feature analysis,model acceleration and data scale,and it will be better applied in the field of music information retrieval.