Remap: recursive estimation and maximization of a posteriori probabilities in transition-based speech recognition