Towards better GMM-based acoustic modeling for spoken language identification

Gaussian Mixture Model (GMM) is a widely used, simple and effective modeling approach for spoken language identification. Traditionally EM algorithm is used to train this model. In this paper we propose a new method named WA-GMM (Weight Adapted GMM) for estimating the weights of GMM Gaussian components using bag-of-unigram and Support Vector Machine (SVM): SVM weights which are trained on bag-of-unigram vectors, are used as new weights for GMM Gaussian components. These new weights act better than the weights resulted by EM algorithm. Our experiments on 3 different LID systems on 4 languages from OGI-TS multi-language corpus prove our claim.