Speaker adaptation using improved MAP estimation with small amount of adaptation data

In this study, we focus on the small vocabulary isolated word recognition system for a portable device such as the remote controller. MAP estimation is a well-known and reliable speaker adaptation technique based on Bayes theory. However, it is difficult to estimate the parameters without the corresponding adaptation data. In this report, we propose a method which solves this problem of MAP estimation. This method is an efficient approach for a small amount of adaptation data. It uses the similarity between states of acoustic models measured by Bhattacharyya distance, and estimates all parameters without the corresponding adaptation data. In experiments of a speaker dependent word recognition using a database consisting of 100 Japanese city names, the proposed method achieved 78.7% recognition accuracy compared to 77.8% of the conventional MAP estimation when 10 adaptation words were provided by target user.