Posterior weights and Gaussian selection for spoken language recognition

This paper investigates the use of the posterior weights of GMMs for spoken language recognition, the goal of which is to determine the language spoken in speech utterances. Since the modeling of distribution is based on the component weights, the number of components in the GMM has to be sufficiently large so as to provide enough degree of freedom. To this end, Gaussian selection technique is applied to speed up the run-time computation. Another problem in using posterior weights is the distance metric. We treat the posterior weights as the probability mass function of a discrete random variable, for which rigorous similarity measures can be easily defined. The proposed language recognition system achieves state-of-the-art performance on the 1996, 2003, 2005 and 2007 National Institute of Standards and Technology (NIST) language recognition tasks.