Research Of Music Chord Recognition Based On Support Vector Machine

With the rapid development of music information retrieval, the traditional musicretrieval method based on the key words limits the application of cover songidentification more and more, many researchers begin to study the new contend-basedmusic retrieval technology. Chord is a typical mid-level feature that it can fullydemonstrate various properties of music. Chord recognition is the base of automaticmusic label, which plays an important role in the fields of song cover recognition,audio segmentation and audio matching etc.The thesis combines the related knowledge of music theory, signal processing andpattern recognition and introduces an algorithm of chord recognition based onSupport Vector Machine and improved Pitch Class Profile feature. The paper includesthe following aspects:At first, the thesis proposes a new feature extraction algorithm which combines thePitch-frequency Cepstral Coefficients(PFCC) with Pitch Class Profile(PCP) since therecognition rate of the same chord is low among the different instruments. Accordingto the music theory knowledge, the changes of chord often occur at the beat. So wefirst extract the beat information of the signal by using the dynamic programmingalgorithm and then extract the feature based on beat tracking instead of the methodbased on frame.Secondly, the thesis introduces the different effects of the different pitch in thesame beat and the effects of different beats on the chord recognition.At last, the thesis describes the chord recognition system based on SVM detailly.The total number of chord type is24which is based on major triad and minor triad.We use the improved PCP as the new chord recognition feature and realize the chordtranscription and recognition by SVM method. The results show that the ratios ofchord recognition have increased2.5%~6.7%after using the improved PCP thanusing the traditional PCP.