Research On Feature Fusion Based On TF-IDF Of Audio And Lyric For Music Eotion Detecion

Music is now acting as the mainstream of multimedia resources and plays a veryimport role in human daily life, since the information and multimedia technologydevelop rapidly. As one of the important communication media of human, music carriesmuch sentiment information. It has important significance to promote the developmentof the natural Human-Computer interaction and multimedia technology to detect theemotion which is the connonation of music with the help of computer. This essayconsists of the flowing major aspects:①Firstly it conducted emotion dection of music solely based on lyric features. Itbases on text categorization to complete the mission. In this part, the CHI featureselection method is analysed and an impoved CHI approach is proposed. The newapprocach makes use of the difference of the different CHI values of the same term todifferent categories to select more discriminative features. Then SVM classifier isconstructed and the comparation of two CHI methods is presented. The experimentsverify the new presented method.②Secondly it fuses lyric features and audio features to classify the musicaccording the emtion included in the music. A feature fusion model based on TF-IDF ispresented. At first, low audio features are extracted. Secondly, by means of LBG-VQalgorithm, the low audio features are mapped into acoustic words to generate the bag ofwords model of audio data. At last, acoustic words and text words are concatenated intovector space to generate the Term-Document matrix of songs. In this process, TF-IDFfunction is used to calculate the term weight. The conventional method which straightconcatenates audio low features and lyric features and the presented approcach arecompared. The result of experiment demonstrates the new method’s effectiveness.