Two-step Semi-supervised Approach for Music Structural Classification

Structural segmentation involves finding hoogeneous sections appearing in a song. The name of these sections depends on the genre of interest. The task is particularly challenging as each of the misclassifed labels gives spurious change points. We have proposed a two step method by finding the boudaries present in the song followed by segment labelling. Novel idea of transforming the features into posteriori space using unsupervised model fitting is elucidated along with interesting behaviour of increasing the mixtures in GMM. Advanced techniques such as hidden markov model is also shown to give reasonable results for the task of boundaries detection. Finally the task of labelling the segments within the detected boundaries are carried out unlike the existing methods which give labels such as A, B, and C. The results are presented for boundary and segment level evaluation.