Anja Belz, SRI Cambridge
Multi-Syllable Phonotactic Models

In this talk I will describe an approach to constructing phonotactic models for use e.g. as linguistic knowledge components in speech recognition. The underlying theoretical approach is based on the multi-syllable hypothesis which holds that syllable phonotactics vary considerably in correlation with a range of factors that include, but are not limited to, a syllable's position within a word and relative to primary stress, and that such variations cannot adequately be captured by standard single-syllable approaches. I will describe a finite-state formalism that facilitates automatic model construction and generalisation, and show how it can be used to construct multi-syllable phonotactic models from data sets of German, English and Dutch phonological words.