SVM binary decision tree architecture for multi-class audio classification

The paper presents the support vector machine binary decision tree scheme (SVM-BDT) used for broadcast news (BN) audio classification. The SVM-BDT architecture was designed to solve multi-class discrimination problem of considered acoustic events: pure speech, speech with music, speech with environment sound, music, and environment sound. Its performance was investigated by using Mel-frequency cepstral coefficients (MFCCs), as a powerful signal parameterization technique, for each SVM binary classifier. The one-against-all strategy in combination with Euclidean distance algorithm was implemented in discrimination process, in order to decrease the influence of missclassification between each class.