| The increasing amount of digital music content has sparked a widespread interest in the area of MIR (Music Information Retrieval) such as automatic playlist generation, music recommendation systems and automatic genre classification. There is a need for developing faster processing systems while maintaining very good accuracy. This work explores semi-supervised learning [17] based on manifolds in relation to automatic genre classification.;A typical music file or its feature representation has thousands of samples. The ability of humans to identify music within a few seconds means that the intrinsic dimensionality of music must be much lower than the actual number of samples, and this dimensionality in fact depends on the number of degrees of freedom of musical instruments. Manifold learning tries to find a low dimensional structure in the music features thus speeding up computations. Laplacian Eigenmaps are used for dimensionality reduction of music features. A combined music feature representation comprising of different fractions of timbre, rhythm and melody features is used. Melody feature, based on chroma and originally used to differentiate between cover songs, has been extended for use in genre classification since it can efficiently capture harmonic patterns in music. Euclidean and Cosine distance measures between the feature representations of tracks are used to obtain the similarity matrix. Regularized Least squares and Support vector machines are used for classification. One against all technique, a simple and efficient technique is used for multi class classification.;A novel multi stage classification approach is proposed. At each stage of classification, a different combination of music features is used in order to exploit the genre - music feature dependency which gives good classification accuracy. A good distribution of music in both training and testing yields almost zero classification error. Experiments conducted in this work also demonstrate the benefits of using unlabeled examples in addition to labeled examples. Random selection of songs for training and testing together with random assignment of labels makes the scheme more robust in learning the classifier. The low dimensional manifold approach preserves the structure in the data and helps in speeding up the classification process significantly. |
