Learning the Model Structure of Dynamic Bayesian Networks for Automated Speech Recognition