Audio classification based on SVM-UBM

Audio classification is an important issue in current audio processing and content analysis researches. In this paper we present a high-accuracy audio classification algorithm based on SVM-UBM using MFCCs as classification features. Firstly MFCCs are extracted in frame level, then a Universal Background Gaussian Mixture Model (UBM) is employed to integrate these sequences of frame-level MFCCs within a clip to form the clip-level feature, finally audio classification is performed using SVM with these clip-level features. Four audio types are considered: speech, music, speech over music and environmental sound. The experimental results show that our classification algorithm performs superior to other SVM-based classification system using traditional clip-level features.