Performance Through Consistency: MS-TDNN's for Large Vocabulary Continuous Speech Recognition