Content-based audio classification using support vector machines and independent component analysis

In this paper, we present a new audio classification system. First, a frame-based multiclass support vector machine (SVM) for audio classification is proposed. The accuracy rate has significant improvements over conventional file-based SVM audio classifier. In feature selection, this study transforms the log powers of the critical-band filters based on independent component analysis (ICA). This new audio feature is combined with mel-frequency cepstral coefficients (MFCCs) and five perceptual features to form an audio feature set. The superiority of the proposed system has been demonstrated via a 15-class sound database with a 91.7% accuracy rate.