Research On The Music Classification Method Based On Correlated Topic Model

With the emergence of digital music on the Internet, automating access to musicinformation through the use of computers has intrigued music fans, computer scientists,engineers, librarians, musicologists, cognitive scientists, music psychologists, businessmanagers and so on. However, current methods and techniques for building real-worldmusic information retrieval systems are far from satisfactory. One of the key problems ismusic automatic classification. As an important part of music information retrieval, musicclassification has become a popular research topic.Since the90s of last century, the study in this field has made significant progress andgot some important achievements, some pattern recognition method such as support vectormachine(SVM),hidden markov model(HMM), neural network(NN) have been successfullyused in music classification and made some successful results. However, a large number ofhigh-dimensional training examples must be collected to train those classification systems.This is a labor-intensive and error-prone process.Recently, some researchers have applied LDA(Latent Dirichlet Allocation) model inaudio information domain. However, LDA can’t control the correlation between topics. Inorder to overcome this shortcoming, Beli et al. proposed correlated topic model (CTM).Based on this, in this thesis, music classification method based on CTM was done researchdeeply to solve these problems, the main research results are concluded as follows:1. Analyzed and summarized the role problems and the advantages anddisadvantages of exiting methods used in music classification;2. Proposed music classification method combing CTM and HMM, in order toimprove the overall performance of music classification system. With the aim of making the CTM model function effectively, density-based spatial clustering of applications withnoise (DBSCAN) is used to determine the topic number of CTM;3. Proposed a new model called Dynamic Correlated Topic Model (DCTM), whichimproved CTM and made it has dynamic. Combing DCTM and HMM as musicclassification method which used the dynamic modeling of DCTM, to improve theperformance of music classification system;4. We used public dataset to verified the effectiveness of the provided method andbuilt music classification experimental system based on the proposed theoretical method;Finally, the research work involved in the thesis was summarized and the futuredevelopments were forecasted.