A study on speaker adaptation of large vocabulary

Abstract

In this paper, the authors propose a speaker adaptation algorithm. There can be a difference of recognition result by a speaker's characteristics although a speaker independent system has a overall good performance. MAP (maximum a posterior) formulation is developed to adapt the characteristics of a speaker with estimation of the HMM (hidden Markov model) parameters from the training data. The proposed adaptation algorithm is evaluated in a large-vocabulary continuous speech recognition. In the experiment, the authors compare the recognition accuracy of the adapted acoustic models. In the experimental results, the MAP algorithm achieves up to about 40% additional reduction of error in phoneme recognition.