Syllable-length path mixture hidden Markov models with trajectory clustering for continuous speech recognition

Recent research suggests that modeling coarticulation in speech is more appropriate at the syllable level. However, due to a number of additional factors that can affect the way syllables are articulated, creating multiple acoustic models per syllable might be necessary. Our previous research on longer-length multi-path models has proved that data-driven trajectory clustering to be an attractive approach to derive multi-path models. However, the use of single distribution with unvarying covariance to model a trajectory cluster may degrade its capability of detecting pronunciation variants. In this paper, we propose a new method, namely path mixture hidden Markov model, to alleviate the adverse effects of trajectory clustering. The improvement on performance observed in continuous speech recognition experiments show path mixture model is a very effective approach.