Inducing Probabilistic Syllable Classes Using Multivariate Clustering -gold

An approach to automatic detection of syllable structure is presented. We demonstrate a novel application of EM-based clustering to multivariate data, exempliied by the induction of 3-and 5-dimensional probabilistic syllable classes. The 3-dimensional models were subjected to a pseudo-disambiguation task, the result of which shows that the onset is the most variable, or least predictable, part of the syllable. An extensive qualitative evaluation shows that the method yields phonologically meaningful syllable classes. We then propose a novel approach to grapheme-to-phoneme conversion and show that syllable structure represents valuable information for pronunciation systems.