Improving Performance of Multiclass Audio Classification Using SVM

Audio classification has found widespread use in many emerging applications. It involves extraction of vital temporal, spectral and statistical features, and using these in creating an efficient classifier. Most of the audio classification work has been done on binary class classification. In our work we suggest best suited features for classification of different audio classes. Here, we present an algorithm for audio classification that is capable of segmenting and classifying an audio stream into speech male, speech female, music, noise and silence. The speech clips are further segment into voiced and unvoiced frames. A number of timbre features have been discussed, which distinguish the different audio formats. For pre classification, Probability Density Function (PDF), which is a threshold-based method, is performed over each audio clip. For further classification, K-Nearest Neighbor (KNN) and Support Vector Machine (SVM) Classifiers are proposed. Experiments have been performed to determine the best features of each binary class. Utilization of these features in multiclass classification yielded accuracy 96.34% in audio discrimination.