An abductive view of high level speech recognition

Here we assume that a phonetic representation of spoken utterances is somehow available, and focus on the problem of identifying the exact words associated with the utterances. This problem may be viewed as explaining the given sequence of phonemes in the utterance using an intended sequence of words. The problem of finding the best explanation for some observed data has been called abductive inference in the artificial intelligence community. Human speech is rife with ambiguity. Ambiguity is a hallmark of abductive tasks. Homophones, or words that have different meanings but the same pronunciation, and continuous speech pose problems to automatic speech recognition. Parsimonious covering has been used to model abductive inference in a variety of task domains, including the problem of understanding written text. We have been studying if similar techniques may be applicable to resolving the ambiguities in high level speech recognition and present the work in progress here. When phonetic-lexical knowledge, such as what is available in a dictionary, is used as background knowledge, it appears that abductive techniques may be applicable to speech.<
>