A study on the use of bi-directional contextual dependence in Markov Random Field-based Acoustic Modelling for speech recognition

Abstract

In this paper, by using the formulation of the missing-data problem, a general framework for statistical acoustic modelling of speech is presented. With the motivation of utilizing bi-directional contextual dependence in acoustic modelling, a bi-directional hidden Markov modelling approach for speech recognition is studied and the importance of the bi-directional contextual dependence for speech recognition is identified by a series of comparative experiments. Furthermore, hidden Markov random field (MRF)-based acoustic modelling techniques using our previously proposed contextual vector quantization (CVQ) method and iterated conditional modes (ICM) algorithm, which is very suitable for parallel processing implementation, are also attempted. Their viability is confirmed by a series of preliminary experiments in a speaker-independent isolated English letter recognition task.