Research And Application Of Musical Emotion Parameterized System

With the development of the Internet, the number of digital music increases so fast that anew method for classify and manage music urgent need. In recent years, domestic and foreignscholars engage in extensive and in-depth research for music retrieval, but fail to obtain awide range of applications. On the one hand, music retrieval is a multidisciplinary field, it isdifficult to research; on the other hand, most recent researches classify and manage music bygenres or emotional labels, just like traditional way, so this approach would be limiting.Considering the disadvantages of music retrieval researches based on musical emotion,this paper proposes a parameterized method of musical emotion for music retrieval. Thismethod extracts the emotion features of music, composites a feature vector, then compressesdimensions using the Fisher algorithm, finally obtains rhythm, tone and timbre threeparameters by training. The bigger parameter reflects the stronger musical emotion. Theresearch achievements are mainly as followed:Firstly, the research of emotion features shows that MFCC is a very important parameter,it largely determine the classified accuracy of musical emotion. The research of MFCCdimension number shows that13or14is reasonable. All different features are not mutuallyexclusive, but complement each other, so using more features can improve the classifiedaccuracy.Secondly, the compare experiments of Fisher and SVM algorithm show differentclassified performance. When in few categories of musical emotion, such as two, Fisheralgorithm should be the first choice in order to conveniently design classifier and conservecomputing resources. When in many categories, SVM algorithm should be the first choice tomaintain high classified accuracy. When in very much categories, it is the selection ofemotion features, not classifier algorithm, plays a key role, more researches should be spendon this aspect.Finally, a musical emotion parameterized system is designed. The system framework isdata-flow model based on Marsyas library and Fisher algorithm is chose as classifier. Rhythm,tone and timbre three parameters are obtained by training through a lot of music samples, thenthese parameters are normalized as0to1. At last, the application of musical emotionparameterized system is complete.The result of testing experiments show that the parameterized system this paper proposesreach88%classified accuracy, basically meets the actual application requirements, providessearch engine for the music management software, promotes the development of musicretrieve technology.