Comparative Research On Song Similarity Via Deep Learning

In recent years,along with lots of excellent original songs released on the popular music platforms,the similarity even plagiarism between songs have been paid more and more attention to.The results produced by the research on song similarity can be used for the plagiarism judgment of songs,also can be used to prevent the song works of the creator from being similar to existing songs.Therefore,it is of great significance to carry out corresponding research on song similarity.At present,in the research field of song similarity,researchers use multiple methods to realize the extraction of song features and then use multiple algorithms to compare the various types of song features.Combined with machine learning technology which holds the powerful feature extraction capabilities of data,good results have been achieved in the comparison of song similarity,such as listening to songs and humming retrieval.However,for the most of current research,people mainly focus on the entire songs,a lot of research is left for the comparison of the similarity between the main song fragments.In this thesis,a comparative research is conducted on the song similarity which is based on the fusion features for song fragments.The main work and results are as follows:1.A model named CDmax_SVM is constructed to compare the fusion features of song fragments.In the model,Cosine Distance(CD)and Dmax algorithm for cover recognition are introduced.The two features,instantaneous frequency Pitch Class Profile(PCP)and Mel-Frequency Cepstral Coefficients(MFCC),are extracted from the samples of test songs and songs in the music library.Four similarity values are obtained by calculating the Dmax value and cosine distance of each feature between two songs.A four-dimensional fusion feature vector is composed of the four similarity values and input into the Support Vector Machine(SVM).The linear kernel function is selected to determine whether the two songs corresponding to the feature vector are similar.2.The song fragments are edited and combined into similar and dissimilar combinations.The performance of the model is verified by extracting the feature vectors of the songs between the combinations,constructing a training set and a test set.Dmax and CD are combined with SVM to form models named Dmax SVM and CD SVM.The performance of the two models is compared with CDmax_SVM.From the overall result trend,the training accuracy of CDmax_SVM fluctuates in the range of 90.1%～92.3%,and the test accuracy fluctuates in the range of 87.6%～93.3%,showing better performance than Dmax SVM and CD_SVM.The effectiveness of CDmax_SVM is proved by the results.3.The covers80 data set is used for similar song search experiments,and the results are compared with those of Dmax ranking and cosine distance ranking.The original singing song fragments are used as the test samples,and the corresponding fragments of the cover songs and the previously used song fragments are used to form a test music library.The similar song search results obtained by the CDmax_SVM are sorted by the size of the Dmax value calculated by the PCP feature,and compared with the sorting results of the Dmax and cosine distance calculated based on the PCP feature.The TOP-3 index of the CDmax_SVM combined with the Dmax sorting is 42,the TOP-3 index of Dmax sorting is 38,and the TOP-3 index of cosine distance sorting is 12.The accuracy of CDmax_SVM model is proved to be a little better than Dmax and CD in finding similar songs.