Improving Melody Classification by Discriminant Feature Extraction and Fusion

We propose a general approach to discriminant feature extraction and fusion, built on an optimal feature transformation for discriminant analysis [6]. Our experiments indicate that our approach can dramatically reduce the dimensionality of original feature space whilst improving its discriminant power. Our feature fusion method can be carried out in the reduced lowerdimensional subspace, resulting in a further improvement in accuracy. Our experiments concern the classification of music styles based only on the pitch sequence derived from monophonic melodies.