Syllable-based constraints on properties of English sounds

This thesis outlines a phonological representation and corresponding rule framework for modelling constraints on an utterance's acoustic-phonetic pattern. The proposed representation and framework of rules are based on the syllable and suggested as an alternative to other representations that are primarily segment-based. Specifically, the traditional notion of a segment is abandoned at the systematic-phonetic level of representation and replaced with a description of an utterance in terms of acoustic properties. Acoustic properties are related directly to distinctive features which comprise an utterance's underlying phonemic description. This relation is specified in the form of a grammar. The grammar incorporates realization rules that allow acoustic properties to be described in direct quantitative terms. Further, properties that are associated with bundles of features are not required to temporally align in the sound stream. Constraints on patterns that comprise the acoustic representation of an utterance are stated as conditions on the realization of well-formed syllables, where an immediate constituent grammar is used to represent syllable structure at the phonemic level. The details of the proposed rule framework are provided, as well as an empirical justification for the use of the syllable in the organization of acoustic-phonetic, phonotactic, and lexical constraints. Justification is provided in the form of results of three sets of experiments. In the first set of experiments, the role of the syllable in describing the distributions of the allophones of stop consonants in American English was examined. The hypothesis is that the syllable simplifies this description. Binary regression trees were used to quantify the extent to which knowledge of a stop's syllable position in conjunction with other contextual factors aids in the prediction of its acoustic realization. The principle of Maximum Mutual Information was used to construct trees. In the second experiment, syllabic constraints on the underlying phonemic description were examined. Specifically, mutual information was used to quantify collocational constraints within the syllable. A hierarchical cluster analysis was used to determine a syllable internal structure that is consistent with mutual information statistics gathered from a 5500-token sample of syllables extracted from a 20,000-word dictionary. Finally, in the third set of experiments, a model of lexical representation is proposed and tested. The model is based on the notion of