A study of audio classification on using different feature schemes with three classifiers

Abstract

Audio classification is an important tool used in audio retrieval and other audio processing application areas.In this paper, five kinds of audio data including speech, symphony, jazz, light music and concerto are studied. Audio features are extracted by audio analysis, and formed into different feature sets. The performance on applying Support Vector Machine, Fisher Kernel Classifier and Potential Function Classifier to different audio feature sets is then examined. Finally, this paper presents the results of a number of experiments, which attempt to increase the classification accuracy by combining three classifiers and different feature sets. The classification accuracy can achieve 97% by using a combination of feature set 1, set 2, set 3, and set 5 with Support Vector Machine and Fisher Kernel Classifier, the accuracy is close or better than the ones reported on the similar data sets and using other classifiers. (Abstract)