Modelling the prepausal lengthening effect for speech recognition: a dynamic Bayesian network approach