Music Classification with Partial Selection Based on Confidence Measures
Wei Chai, Barry L. Vercoe · 2005
Music classification is a useful technique that enables automation of labeling musical data for searching and browsing. One method for music classification is to label the sequence based on the labels of individual frames. This paper investigates the performance of using confidence measures to select only the most “useful” frames to make the decision of the whole sequence. Confidence measures for Support Vector Machines (SVM) and Predictive Automatic Relevance Determination by Expectationpropagation (Pred-ARD-EP) are particularly examined. Experimental result shows that selecting frames based on confidence significantly outperform selecting frames randomly and the confidence measures do, to some extent, capture the “usefulness” of musical parts for classification.