Research Of Chord Recognition Based On Auditory Images

With the growing popularity of Internet technology, massive music multimedia information is appeared in network. Therefore, content-based music information retrieval has become an important method for a user to quickly and efficiently retrieve the desired music data. Chord, one of the important middle features of the music signal, is the basis of the music information labels. And chord as the inherent characteristics of the music, it can describe the polyphonic content and chromatic structure. Chords have widely applied in audio detection and segmentation, musical sentiment analysis, audio scores alignment and other fields.The thesis proposes auditory image features of chords, which combines with human auditory perception system, auditory psychology, musical theory and pattern recognition. Then, through sparse representation classification complete chords recognition and classification. The method specific implementation is as follows:Firstly, according to chord unit, cut a integrated music into pieces by using dynamic programming algorithm to extract the beat information of the music signal, so that the algorithm is more robustness and not affected by beats. Then use auditory image model, which converts the one-dimensional music signals into two-dimensional auditory images, through pre-cochlear processing, basilar membrane activity, neural activity pattern and strobe temporal integration.Secondly, the thesis extracts fine-structure of different chords auditory images by scale-invariant feature transform(SIFT) and spatial pyramid matching(SPM).Finally, establish the auditory image-based chord recognition system, which divides chords into 24 categories, and for each to build a database, then extracts average of features. Labels, marked by Chris Harte, are used to train a supervised sparse representation classification model. After that, use the model to recognize the test chords.Experimental results show that the proposed algorithm, based on auditory images features, highest correctly chord recognition rate is 76.2%, is also higher than MFCC feature, which is based on the human auditory characteristics.