Context dependent phoneme duration modeling with tree-based state tying

In this paper, we propose phoneme duration modeling methods with tree-based state tying. Two kinds of phone duration modeling methods are suggested. The first is context independent phoneme duration model in which duration parameters are stored in each phone. The second is context dependent duration model in which duration parameters are stored in each state being shared by context dependent phone. We split duration parameters of each context dependent phoneme into three kinds of tied states estimated by tree-based clustering. The duration of each state is modeled by Gamma distribution function. Both HMM and duration parameters are stored in states tied for expressing all context dependent phones in a phone. The duration parameters of context dependent phoneme are automatically generated from state duration parameters in the initialization stage of recognition. Context dependent phoneme duration model is compared with context independent phoneme duration model as well as with no duration model. We make experiments with database collected through telephone network service. Experimental results demonstrate that duration information rejects OOT (out-of-task) words very well and that context dependent duration model yields the best performance among three methods.