Audio Signal Classification Based on Optimal Wavelet and Support Vector Machine

Audio signal classification is done by extracting relevant features from a sound and using these extracted features, it identifies into which of a set of classes the sound is most likely to fit. In this paper, an improved feature vector formation technique for audio classification and categorization is presented. This technique makes use of wavelets to extract the features of audio data. Wavelets are first applied to decompose the signal and to extract acoustical features such as sub-band power, brightness, band width and pitch information. The additional features, such as frequency cepstral coefficients also extracted to accomplish audio classification. This paper also proposes a different frame size computation strategy that uses 512 samples per frame. The overlap size between windowed frames of the proposed feature extraction algorithm is redesigned and a vector formation process is carried out after feature extraction. Finally, the proposed method uses a bottom-up support vector machine over these acoustical features and additional features. The bottom-up support vector machine categorization strategy uses an iterative procedure to match a given audio to progressively larger subsets or categories of classes. In addition, two support vector machine training parameters such as the upper bound and the variance of the exponential radial basis function are modified to improve the accuracy of classification. It is shown in experimental results that the categorization accuracy of a given audio sound can achieve 100% in the top 1 and top 2 levels.