A Bayesian Information Criterion Based Approach for Model Complexity Selection in Speaker Identification

Model complexity selection is important in the task of speaker identification. A Bayesian information criterion (BIC) based approach for model complexity selection is proposed in this paper. The speaker models are trained with the speech features. Then the BIC values of speaker models are calculated. In order to reduce the computation of training speaker models with different complexity, the greedy strategy is used to search the locally optimal model complexity. The experiments compare the model complexity selection effect of the proposed approach to the fixed size fashion and other model selection methods. The results demonstrate the effectiveness of the proposed approach.