UBM data selection for effective speaker modeling

This paper presents a UBM data selection method for robust training. We know that there is no promise that more training data guarantee better results. Therefore, the way of sub-sampling and effective training become important. The proposed method uses the feature vector selection with the maximum-entropy criterion. The maximum-entropy shows the diverse characters of speaker and minimum redundant information as well. The UBM training data is investigated on three datasets to compare the proposed method with the conventional sub-sampling approaches. We conducted experiments on the 2008 NIST Speaker Recognition Evaluation corpus shows that the proposed method outperforms the conventional one in speaker recognition.