Continuous speech recognition using two-level LR parsing