Abstract
In this thesis, we try to use Hidden Markov Model (HMM) to solve the problem of large vocabulary recognition of Taiwanese. The phone units we used are Initial and Final units. The recognition accuracy we achieved in our initial study is 90.6% / 97.4% for Top1 / Top5 candidates respectively. A traditional left-to-right HMM with no skipping states has been suffered from its poor ability to model the duration of each state in a speech unit. Therefore, some implicit or explicit duration models have been proposed in the literatures during the past years. In this thesis, a new hybrid duration modeling technique based on an infinite-duration HMM and a finite-duration HMM is proposed. The application to a pioneer large-vocabulary Taiwanese speech recognition task shows promising results and improvements over the traditional explicit duration modeling techniques.