Modeling cross-morpheme pronunciation variations for korean large vocabulary continuous speech recognition.