MDL-Based Cluster Number Decision Methods for Speaker Clustering and MLLR Adaptation

Speaker clustering is one of the major methods for speaker adaptation. MLLR (Maximum Likelihood Linear Regression) adaptation using transformation matrices corresponding to phone classes/clusters is another useful method especially when the length of utterances for adaptation is limited. In these methods, how to decide the most appropriate number of clusters is an important research issue. This paper proposes to use the MDL (Minimum Description Length) criterion to decide the optimum number of speaker clusters as well as phone clusters according to the size of utterances for clustering and adaptation. Experimental results of speaker clustering and MLLR-based speaker adaptation show that the MDL criterion derives the optimum number of clusters according to the size of utterances, which achieves the highest recognition performance.