Logic-based work and utterance discovery from mixed-detail phonetic lattices

One approach to continuous speech recognition is to segment the acoustic linguistic analysis. It is shown how the insights of A. Smith (1976), K.W. Church (1983), and V.W. Zue (1985) in this regard may be combined in a single system. The algorithms and data structures can be conveniently expressed in a logic-based programming formalism. Input is a phonetic lattice with detail ranging from coarse classification to precise allophonic identification. A four-pass algorithm parses the phonetic input into a word lattice. Closely allied with the algorithm is a data structure for the lexicon. The lexicon's organization promotes locality of reference during the word discovery process. It is shown how a definite-clause grammar parser can be adapted to finding sentences among the words.<
>