Research Of Chord Recognition Based On Sparse Representation Classification

With the rapid development of digital music, the traditional music retrievalsystem which is based on the key words limits the practical applications more andmore. However, the content-based music retrieval without the limitation of keywordsin audio retrieval, but according to the inherent properties of the music, has beenincreasingly concerned about by many researchers. As a typical mid-level feature,chord contains rich information which can represent musical attributes, and has a veryimportant role in analyzing the structure and melody of the music. In this paper,author researches the technology of content-based music retrieval in-depth, andproposes the representation of music chords feature as well as recognition algorithms.This dissertation combines the knowledge of music theory, signal processing andpattern recognitions to propose one chord recognition method which based on sparserepresentation classifier. The main research content of the paper is to create a sparserepresentation model based on the chord as basic unit, and then construct a completechord recognition system based on this model. This paper includes the followingaspects:At first, according to music theory knowledge, we know the changes of chordsoften occur at the beat, so this article proposes a feature extraction method based onbeat detection. According to Daniel Ellis’s method, the input of the entire audio file isdivided into different beats (audio clips), and beats are regarded as the minimum timeinterval of chord changes.Secondly, this paper studied the Pitch Class Profile features that commonly usedin the chord recognition area, and list out the calculation formula. According to latesttheory, this article adopts QPCP as the chord feature, so that a piece of music can berecognized and expressed more effectively with machines.Thirdly, we set up one music chord recognition system based on SRC. In total,the types of chord are24, including major triad and minor triad. For each chord, thedata base was constructed respectively and extracted average feature. Finally, theinput audio file which is regarded as the test sample is identified.At last, we get the supervised SRC by virtue of the label files made by ChrisHarte and compared with classic template-based chord recognition method. Theexperiment results show that the average recognition rate of SRC is78.6%, and is 2.7%higher than the template-based method.