CONSTRUCTING MARKOV MODELS OF WORDS FROM MULTIPLE UTTERANCES

摘要

ABSTRACTThe present invention addresses the problem ofconstructing fenemic baseforms which take intoaccount variations in pronunciation of words fromone utterance thereof to another. Specifically,the invention relates to a method of constructinga fenemic baseform for a word in a vocabulary ofword segments including the steps of: (a)transforming multiple utterances of the word intorespective strings of fenemes; (b) defining a setof fenemic Markov model phone machines; (c)determining the best single phone machine P1 forproducing the multiple feneme strings; (d)determining the best two phone baseform of theform P1P2 or P2P1 for producing the multiplefeneme strings; (e) aligning the best two phonebaseform against each feneme string; (f) splittingeach feneme string into a left portion and a rightportion with the left portion corresponding to thefirst phone machine of the two phone baseform andthe right portion corresponding to the secondphone machine of the two phone baseform; (g)identifying each left portion as a left substringand each right portion as a right substring; (h)processing the set of left substrings and the setof right substrings in the same manner as the setof feneme strings corresponding to the multipleutterances including the further step ofinhibiting further splitting of a substring whenthe single phone baseform thereof has a higherprobability of producing the substring than doesthe best two phone baseform; and (k) concatenatingthe unsplit single phones in an ordercorresponding to the order of the fenemesubstrings to which they correspond.