Research On Autamatic Music Structrue Analysis

Music is now one of the most popular topics to study in audio processing, and it is a good carrier of many kinds of perceptional information. Music structure is one main representation of music piece to organize and express information, and it is one of the key paths to music comprehension. Now it is one of the hottest topics in music information retrieval field.In this thesis we try to segment pop music into intro, verse, chorus, bridge and outro from two aspects of view. One is to detect vocal part in pop music; the other is to analysis pop music structure based on similarity matrix. Since it is that the whole pop music is almost pure instrumental except verse and chorus, so detection of vocal part in pop music can do great help to pop music structure analysis. To solve the problem of detection of vocal part in pop music, we propose MFCC as features, while SVM to classify. Due to the continuity of audio signal, we use low-pass filtering to classification results in the post-processing, and the classification accuracy increase 11.9% in average. Experiment results shows, we can get a quite promising classification accuracy of 85.76% when processing frame by frame. We converts one-dimensional pop music audio to two-dimensional image by using similarity matrix, due to the fact that verse and chorus part always appear more than once in pop music, these parts are presented in the similarity matrix as straight line, so we turn the problem to straight line detection in images. Here we pose the problem of detection straight lines in gray-scale images as an inverse problem. This formulation is based on the use of the inverse Radon operator. The advantage of this formulation is that we can then study the problem within a regularization framework and enhance the performance of the Radon-based line detector through the incorporation of prior information in the form of regularization. Thus we fix intro, bridge and outro parts by detection of vocal part in pop music, determine verse and chorus parts through similarity matrix, and finally we combine the two results to the final structure analysis result.Chorus detection is a hot topic in MIR(music information retrieval), this thesis not only detects the position of chorus, but also gives out other representative period.