Speaker adaptation using Maximum Likelihood General Regression

In this paper, a new method called Maximum Likelihood General Regression (MLGR) is introduced for speaker adaptation. Gaussian means of a speaker independent (SI) model are adapted to the data of a new speaker by assuming a non-linear mapping from the SI Gaussian means to the adapted Gaussian means. MLGR performs a non-linear regression between ML estimates of the means and the SI means using General Regression Neural Network. The proposed method is evaluated on the Wall Street Journal database. Evaluation results show that the suggested scheme outperforms different conventional approaches in the case of short adaptation utterances. We also mathematically prove that the Gaussian means of the adapted model using the MLGR converges to their ML estimates in the case of long adaptation utterances.