A Sample and Feature Selection Scheme for GMM-SVM Based Language Recognition

Discriminative training for language recognition has been a key tool for improving system performance. SVM-based algorithms (i.e. GMM-SVM, GLDS-SVM etc.) are important ones for language recognition. The core of these algorithms is to construct the kernel for comparing the similarity of two sequences. It is known that the mismatch between training and test condition will degrade the performance. In this paper, we proposed a novel sample and feature selection scheme under the GMM-SVM framework, which aims at alleviating the duration mismatch problem. The proposed method is evaluated on NIST 03 and 07 language recognition evaluation tasks with improvement over prior techniques.