Classification of Voice Content in Public Radio Broadcasting Context

Abstract

Mass media has acquired a global character because of the rapid development of information technology. This technological advancement undoubtedly impacts the traditional mass media such as newspapers, television broadcasting, and radio broadcasting where the important changes have occurred in its production and distribution chains. With the evolution of mass media technology, content analysis of radio broadcasting emerged as a major research area which facilitates to automate the radio broadcasting monitoring process. This dissertation focuses on the problem of automating the radio broadcasting monitoring process in Sri Lanka. A proper content classification is required to monitor radio broadcasting content automatically. In this research, more attention goes to the voice dominant content classification of radio broadcasting by employing a multi-class Support Vector Machine(SVM). Multi-class SVM implements as a compound of binary SVM classifiers. This study comparably investigates the performance of “One Vs. One” and “One Vs. All” methods which are known as two conventional ways to build multi-class SVM. One of the most substantial measures in creating such classification is selecting the optimal feature sets for each binary SVM classifier independently. For that, time domain features, frequency domain features, cepstral features, and chroma features are manually analyzed. The two multi-class SVM models are trained based on the selected features. These models are capable of classifying five voice dominant classes such as news, conversations, advertisements without jingles, radio drama and religious programs with accuracies of 85% and 83% respectively for “One Vs. One” and “One Vs. All” models. Therefore “One Vs. One” model is selected as the soundest multi-class SVM classifier for this study