Enhanced polyphonic music genre classification using high level features

The task of classifying the genre of polyphonic music signals is traditionally done using only low level features of the signal. It has been suggested in the literature that high level musical features are also good source of information on music genre. In this thesis high level features have been applied to improve and enhance the task of music genre classification. Features that capture high level conceptual harmonic, pitch, and rhythmic contents of the music are proposed to be used in conjunction with low level features to increase the classification accuracy of polyphonic music signal genre classification techniques. In this thesis chord, chord progression, and beat features are used in conjunction with timbral features to improve the music genre classification methods. Since chord and chord progressions are perceptual concepts and differ from timbral and beat features in nature, they cannot be directly integrated with each other. Therefore specific techniques have been developed in this thesis to capture the high level information of chord and chord progressions into feature vectors so they can be integrated with beat and timbral features. To capture chord information, a statistical chord feature vector is proposed. And to capture chord progression information, a technique called Chord Mining is developed. In this technique, chord progression content of the songs are manifested in genre probability descriptors calculated using a pattern matching algorithm. Our proposed method provides an improvement of 12.4% in the classification results over a commonly compared technique.