Product HMM-based training method for acoustic model with multiple-size units

Abstract

Multiple-size units-based acoustic modeling has been proposed for large vocabulary speech recognition system to improve the recognition accuracy with limited training data. By introducing a limited number of long-size units into unit set, this modeling scheme can make better acoustic model precision than complete short-size unit modeling without losing model trainability. However, such a multiple-size unit acoustic modeling paradigm does not always bring reliable improvement on recognition performance, since when a large number of long-size units are added in, the amount of training data for short-size units will decrease and result in insufficiently trained models. In this paper, a modified Baum-Welch training method is proposed, which uses product hidden Markov models (PHMMs) to couple units with different sizes and enables them to share same portions of training data. The validity of proposed method is proved by experiment results.