A FRAMEWORK AND TOKEN PASSING MODEL FOR CONTINUOUS SPEECH RECOGNITION WITH DYNAMIC BAYESIAN NETWORKS

摘要

Hidden Markov models (HMMs) are the most commonly used stochastic model encoding acoustic features in speech recognition. The token passing model is an abstract model for HMM-based continuous speech recognition to uncouple acoustic models (HMMs) and the language model. Recently, there has been an increasing interest in a general class of probabilistic models: dynamic Bayesian networks (DBNs). Although a huge success of the introduction of DBNs into speech recognition in many areas, the frameworks and recognition algorithms for DBN-based continuous speech recognition are not as mature and flexible as those for HMM-based one. This paper is trying to propose a general framework to inherit most features of state-of-the-art HMM-based frameworks for continuous speech recognition and incorporate the interpretability, factorization and extensibility of DBNs into our framework. The token passing model is adapted for DBN-based continuous speech recognition to achieve this goal and a novel recognition algorithm independent of the upper-layer language model is proposed in this paper.