A new approach for spoken language identification based on sequence kernel SVMs

A new back-end classifier for GMM-LM based language identification systems is proposed in this paper. The proposed system consists of a mapping matrix and a back-end classifier of SVMs as its main parts, located in series after the GMM-LM system. While the mapping matrix maps the language model's output vectors to a new space in which the languages are more separable than before, each SVM in the SVM bank-end classifier separates one language from the others. A new sequence kernel is used for each SVM in the bank-end classifier. As a final stage, a fusion block carries out the task of fusing the SVM bank-end scores with those of the GMM-based LID to achieve higher accuracies. We show that not only our new sequence kernel-based SVMs separate languages more efficiently than common Gaussian mixture and GLDS SVM back-end classifiers, but also our new mapping matrix outperforms common linear discriminant matrix in separating classes from each other and finally the introduction of fusion block leads to even superior performance. The overall accuracy of the LID is noticeably increased in comparison with the other LDA-GMM and LDAGLDS SVM back-end classifiers. Our experiments on 5 languages from OGI-TS multilanguage task prove our claim.