Automatic sound classification of radio broadcast news

Automatic extraction of the index of broadcast streams from radio and television has become a challenging research topic over the last years. The automatic classification of audio types, such as speech, music, noises/atypical events etc, has found numerous applications. In this paper we study the evaluation of different machine learning algorithms, which have successfully been used in other classification tasks, on the task of classification of audio broadcast news. The audio classification scheme consists of pre-processing, audio parameterization with established audio features, and classification to acoustic events. The experimental evaluation was carried out using the Voice of America broadcast recordings database for the Greek language. The experimental results indicated that the best performance, approximately 92% of accuracy, was achieved by the classification scheme using the boosting technique with decision trees.