Musical Genre Classification Using Ensemble of Classifiers

Abstract

The increase of the music databases on the personal collection and the Internet have brought a great demand for music information retrieval, and especially automatic musical genre classi cation. Most automatic music genre classi cation researches have been focusing on combining information from di erent sources than the musical signal. This thesis presents a novel ensemble approach for the automatic music genre classi cation problem using low-level characteristics of audio signals into high-level hierarchically organized genre taxonomies. The proposed approach uses two feature vectors, Support vector machine classi- er with polynomial kernel function and a pattern recognition ensemble approach. The short term low-level audio features are derived from 30ms audio signal frames with a hop-size of 10ms and all the long term low-level audio features have a frame size of 10 seconds. More speci cally, two types of feature vectors for representing frequency domain, temporal domain and cepstral domain short term based audio features and modulation frequency domain long term based audio features are proposed for individual classi cation. For feature selection purposes, we used wrapper method for short term based feature vector and ltering method for long term based feature vector. The support vector machine classi er with polynomial kernel function is employed as the base classi ers for each individual feature vectors. Using our proposed features SVM act as a strong base learner in AdaBoost, so its performance of the SVM classi er cannot improve using boosting methods. The nal genre classi cation is obtained from the set of individual results according to a weighting combination late fusion method and it outperformed the trained fusion method. Music genre classi cation accuracy of 78% and 81% is reported on the GTZAN dataset over the ten musical genres and the ISMIR2004 genre dataset over the six musical genres, respectively. We observed higher classi cation accuracies with the ensembles, than with the individual classi ers and improvements of the performances on the GTZAN genre dataset are three percent on average. This ensemble approach shows that it is possible to improve the classi cation accuracy by using di erent types of domain based audio features.