Research On Related Algorithms Of Automatic Classification Of Music Signals

Nowadays, the amount of digital music is rocketing. Because of complexity of music by nature, poorly-defined conception of sort and limited knowledge of perceptual feature of human hearing, researches on the related topics of music automatic classification are almost at a phase of starting point.Under these circumstances, the speed and efficiency of music resources retrieval depends on music automatic classification, and the potential demands are huge, so it is worth investigating.In order to overcome the limitation of the traditional audio signal analysis theories such as Mel Freqency Cepstrum Coefficients and Time Frequency Analysis in the features extraction of music with plentiful structures and information, this dissertation presented several algorithms reflecting features of music signal more precisely by combining auditory model of human ear, perceptual feature of human hearing, sparse coding, homomorphic analysis, time-frequency analysis and the nature of music signal. And the features of music signals were researched thoroughly.First, All Information MFCC was presented. It arose from the modification of the traditional melody frequency cepstrum coefficients in homomorphic analysis. All kinds of possible combinations of music samples were considered in order to depress spectrum leakage from truncation of signal, enhance actual spectrum of music signal and get the information of change of note.Second, Music Bionic Wavelet Transformation was put forward. It was based on modified Giguere's hearing perceptual model of human ear and the property of quality factor of human hearing system, combining critical bands of hearing, and parameters reflecting dynamic characteristic of human hearing and hearing masking effect were introduced into the wavelet transformation. Therefore, it had two-dimension-independently-adjustable resolution in time and frequency and can be adjusted adaptively by instantaneous amplitude and its first-order differential of music signal.Third, algorithm of feature vectors construction and sparse representation of music was advanced on the basis of maximizing approximate negative entropy of characteristic component of music. It was according to sparse coding and sparselized interpretation of information processed by biology system in neurophysiology, which fully reflects the behaviors and structures of music. Furthermore, with quasi-uncorrelated sparse decomposition, the algorithm was extended to the space of overcomplete feature vectors with more structures and features.Finally, blind linear-mixed musical instruments separation algorithm in noise was proposed. The algorithm utilized the features of time-related statistical characteristics and sequential structures of music signal to realize the successful separation of music signals from different kinds of musical instruments by combining time-delayed covariance matrix and fourth-order cumulant matrix, and making use of robust orthogonal method with average feature structures of some data matrixes.The presentation of the algorithms mentioned above improved the representation of structures and features of music, so the musical instruments can be separated correctly, and with unsupervised clustering methods, music automatic classification was realized.