Incremental Learning Based on Probabilistic SVM and SVDD and Its Application to Acoustic Signal Recognition

To increase the efficiency and accuracy of multiclass acoustic signal recognition, an incremental learning algorithm based on probabilistic support vector data description (SVDD) and support vector machine (SVM) is proposed in this paper. A general classification structure with probability distribution is constructed by applying probabilistic SVDD and SVM to separately train models. Avoiding repeated training, class-incremental learning (CIL) is based on the combination of separately trained models and the overall model construction by kernel density and probability comparison. For its specific application to acoustic signals, Mel-frequency cepstrum coefficient (MFCC) is extracted from the signals and is input to learning algorithms as features of acoustic signals. Experiment results show that the proposed method can applied into acoustic recognition and improve the accuracy compared to other CIL methods.