Automatic selection of acoustic features using a lazy spitting method

Abstract

The increasing amount of music data approaching the scale of ten million of tracks poses the challenge of organizing such huge information. Audio Tag Classification is a sub-area in Music Information Retrieval. Its objective is to predict human motivated tags given the acoustic data. One major problem in this procedure is the training of the classifier. An important step in the training is the selection of the appropriate acoustical features. This paper explores two selection approaches: greedy and spitting. Experimental results indicate that the proposed spitting algorithm has a superior performance both in classification (F-measure score) and speed (lower computational requirements).