On The Multi-class Audio Classification Based On The Support Vector Machines

Support vector machine (SVM) is a statistic learning method based on lesssamples proposed in recent years. In the paper, an AB(Augment Binary)strategyis used to deal with the multi-class audio classification problem. In order to explorethe potential of the SVM in audio classification, the experiment is based on thecommon audio database, which is classified into 13 classes consisting of about 327sounds. The comparison is made between the AB with other popular approaches.The main novel points in this paper are listed following: Firstly, Low Energy Ratioand AMDF are modified during training the samples. In the experiment, theMLER(Modified Low Energy Ratio) and DAMDF achieve better performance.Secondly, a new method for multi-class is introduced to design the audio classifier .The experiential results show that the new method exhibits the better property.Furthermore, the method is the mapping of the multi-class problem to a bi-class one,which allows us to suggest a method for estimating the generalization error by usingdata-dependent error bounds. Finally, in the training phase, a new preprocessingmethod is proposed, in which wild-points are removed with density method and thenvector quantization is made. Such method greatly reduced the dataset in thetraining phase. Furthermore, the Reduced-Set-Method is used to simplify the modelwith reduced the number of support vectors. The experimental results shows thatthe reduced set method is considerably effective to improve the function of classifierfor median-scale data-processing.