Layered Markov models: A new architectural approach to automatic speech recognition

This paper presents the theoretical basis of layered Markov models (LMM), which integrate all the knowledge levels commonly used in automatic speech recognition (acoustic, lexical and language levels) in a single model. Each knowledge level is represented by a set of Markov models (or even hidden Markov models) and all these sets are arranged in a layered structure. Given that common supervised training and recognition paradigms can be also expressed as simple Markov models, they can be formalized and integrated into the model as an extra knowledge layer. In addition, it is shown that hidden Markov models (HMM) and newer HMM2 can be considered as particular instances of LMM.