Research On Electronic Music Feature Analysis And Genre Classification

With the rapid popularization of the Internet and the widespread application of audio compression technology,people acquire and enjoy digital music more through online music platforms,meanwhile music creation has become much more easier.But the resource of music libraries has been expanding exponentially.Thus how to effectively manage a large number of music resources has spawned the subject of content-based music information retrieval,of which music genre recognition is an important research.In recent years,with the rapid growth of the music library,the music genre system has become increasingly large.It has been difficult to achieve excellent results using traditional methods to efficiently with genre recognition problems and the big data corresponding to music libraries.Therefore,in the current context,there is an urgent need to develop a new method to improve the efficiency and accuracy of the genre automatic classification system.This study selected the electronic music which people fond of today as the research object,designed an electronic music genre classification method based on the spectral characteristics,improved the system through several experiments,thus completed the research of electronic music genre recognition,and provide new thoughts for the research of music genre recognition classification for other music forms.This research has mainly focused on the following aspects:It summarized various characteristics that are commonly involved in the classification of music genres,demonstrated and analyzed the characteristics and other factors of each characteristic,finally it selected the spectral characteristics as the focus of research.In order to perform spectrum visualization conversion through real music sequence,it generated short-time Fourier spectrum,Mel spectrum and constant Q spectrum,then compared the differences of these spectrum,finally it illustrated with example the relationship of the three kinds of spectrograms generated by music samples and their genres.Considering the excellent performance of the convolutional neural network in image processing,and the problem of network degradation,we designed an electronic music genre classification model with a spectrogram as an input and a residual network as a classifier.Nine sets of experiments were arranged,with a total of 1,000 electronic music samples from two datasets——The Beatport EDM Key Dataset and The Giant Steps + EDM Key Dataset.In order to verify the effect of depth on the classification model,the three spectrums corresponded to the performance of the classification model,and the network designed in this study are compared with traditional machine learning classification methods.It is concluded that the classification system proposed in this study has achieved an average accuracy rate of more than 70%,which performed better than the traditional shallow classification reached the level of mainstream classification systems.The average accuracy is more than 10% higher than traditional methods.The classification method implemented by layer machine learning even has an accuracy rate close to 80% in the classification recognition of several genres.There is still room for improvement in the classification effect achieved by the classification model.Inspired by the relevant research on audio scene detection,the HPSS algorithm is introduced into the spectrum division.Based on the original three spectrum diagrams,nine spectrum diagrams are divided more to reach the purpose of enriching the feature.The nine spectrums retained the characteristics of the original three spectrums and increased the sensitivity of frequency characteristics.Combining the classification results of the residual network,we proposed a variety of integrated learning methods for spectrograms for the first time.The specific steps are as follows.Firstly the system sequentially generates multiple spectrums of the music sequence,and learns the corresponding features of the various spectrums through the residual network,to produce nine classifiers,then as another input of a multilayer perceptron,the output of the nine classifiers can produce another classification result by stacking integration strategy.After verification,the difference between the classification performance of the integrated learning method and the individual residual network method is evaluated.The results shows that the performance of the integrated learning method is better than that of the individual residual network method,achieving an average classification accuracy of 86.17%?...