A method to integrate additional knowledge sources into HMM based on junction tree decomposition

摘要

Most current automatic speech recognition (ASR) systems use statistical data-driven methods based on hidden Markov models (HMMs). Although such approaches have proved to be efficient choices, ASR systems often still perform much worse than human listeners, especially in the presence of unexpected acoustic variability. Only a limited level of success can be achieved, by relying only on statistical models and mostly ignoring the additional knowledge available. We propose a new method of integrating various kinds of additional knowledge sources into an HMM-based statistical acoustic model in this paper. We utilized the junction tree algorithm to achieve efficient integration due to increased model complexity. This is since it facilitates the decomposition of the joint probability density function (PDF) into a linked set of local conditional PDFs. This way, a simplified form of the model could be constructed and reliably estimated using limited training data. We evaluated how efficient the proposed method was on an LVCSR task using two different types of accented English speech data. The experimental results revealed that our method improved word accuracy with respect to the standard HMM.