Classification of Voice Content in the Context of Public Radio Broadcasting

Abstract

With the rapid development of mass media technology, content classification of radio broadcasting has emerged as a major research area facilitating the automation of radio broadcasting monitoring process. This research focuses on the voice dominant content classification of radio broadcasting by employing a multi-class Support Vector Machine (SVM) in order to automate monitoring of radio broadcasting in Sri Lanka. This study investigates the performance of “One Vs. One” and “One Vs. All” methods known to be two conventional ways to build a multi-class SVM. These two multi-class SVM models are trained to classify five voice dominant classes as news, conversations, and advertisements without jingles, radio drama and religious programs. One of the substantial measures in creating such a classification is selection of the optimal feature sets. For that purpose, time domain features, frequency domain features, cepstral features, and chroma features are manually analyzed for each binary SVM classifier independently. Two multi-class SVM models are trained based on the selected features and the “One Vs. One” model was able to better classify the recordings with an 85% accuracy compared to 83% accuracy achieved by “One Vs. All” model. Further, the results revealed the importance of careful feature selection in order to achieve higher classification accuracies.