Bhattacharyya-based GMM-SVM System with Adaptive Relevance Factor for Pair Language Recognition
In this paper, we develop a hybrid system for pair language recognition using Gaussian mixture model (GMM) supervector connecting to support vector machine (SVM). The adaptation of relevance factor in maximum a posteriori (MAP) adaptation of GMM from universal background model (UBM) is studied. In conventional MAP, relevance factor is empirically given by a constant value. It has been proven that the relevance factor can be dependent to the particular application. We use the relevance factor to control the degree of influence from the observed training data for more effectiveness. In order to design a robust pair language recognition system, we develop a hybrid scheme by using separate-training Bhattacharyya-based kernels with the adaptive relevance factor applied. The pair language recognition system is verified on National Institute of Standards and Technology (NIST) language recognition evaluation (LRE) 2011 task. Experiments show the improvement of the performance brought by the proposed scheme.