The effects of equivalence classes on parsing phonemes into words in continuous speech recognition

This study assesses the effect of using different phonological units on parsing a given string of phonemes into words in a continuous speech recognizer. It is shown that when an input utterance is encoded using a representation intermediate between the broad-classes, in Huttenlocher & Zue 1984, and the 44 phonemes of Received Pronunciation, the input can be parsed into more than 10 million different word-strings; and that even when all 44 phonemes are implemented, some input utterances can still be parsed into in excess of 10 000 different word-strings. The results suggest that there is insufficient information in a mid-class representation for post-lexical processing to identify the target word-string.