Generating Gaussian Mixture Models by Model Selection For Speech Recognition

While all modern speech recognition systems use Gaussian mixture models, there is no standard method to determine the number of mixture components. Current choices for mixture component numbers are usually arbitrary with little justification. In this paper we apply some common model selection methods to determine the number of mixture components. We show that they are ill-suited for the speech recognition task because increasing test set data likelihood does not necessarily improve speech recognition performance. We then present a model selection criterion modified to better suit speech recognition, and its positive and negative effects on speech recognition performance.