Audio Event Recognition Using GMM Supervector Based SVM

In this paper, we investigate GMM supervector based Support Vector Machine (SVM) with spectral features for audio event recognition. GMM supervector is obtained by stacking the mean of each Gaussian component that trained by adaptive GMM training. The GMM supervector then is used as input feature for SVM. Experimental results on an audio database demonstrate that our proposed approach significantly outperforms standard GMM-UBM on audio event recognition. Specifically, error rate is decreased from 26.52% to 14.97% by the GMM supervector based SVM using GUMI kernel as compared with GMM-UBM approach for 16 mixtures.