Dissertation/Thesis Abstract

The characterization of phonetic variation in American English schwa using hidden Markov models
by Lilley, Jason, Ph.D., University of Delaware, 2012, 150; 3526450
Abstract (Summary)

The discovery and characterization of a phonetic segment's variants and the prediction of their distribution are two of the chief goals of phonology. In this dissertation, I develop a new, mostly automatic technique for discovering and classifying contextual variation.

The focus is on a set of sounds in English that undergoes considerable phonetic variation, the reduced vowels. These are typically transcribed as a single mid central vowel known as a schwa, but in reality, schwas surface in a number of forms that may be perceptually distinct. Although many studies have demonstrated how the schwa varies in particular controlled contexts, it is not clear which of the contextual effects are strongest, whether there are more undiscovered effects, and what precisely the distribution of schwa variants is. Thus, I use a particular subtype of hidden Markov model, in which states are positioned parallel to each other in order to model alternative pronunciations, so as to examine the variation in American English schwa that is determined by its linguistic context. The technique was used on a corpus of 15 speakers containing over 22,000 schwa tokens.

The results indicate that there is significant variation in both the first and especially the second formants of schwas. High front schwas are more common than low back schwas, although mid central schwas are most common. High back and low front schwas are rare. This variation is mostly accounted for by the schwa's immediate phonetic context—particularly the following segment—but non-phonetic factors, such as the schwa's underlying representation, the surrounding morphological structure, and even part of speech, have an effect. Finally, some of these non-phonetic effects have a small but measurable effect on the intelligibility of synthetic speech in a perception experiment. The results have implications for the automatic phonetic analysis of large corpora, as well as for applications such as automatic speech synthesis.

Indexing (document details)
Advisor: Vogel, Irene, Bunnell, H. Timothy, Adams, Frederic
Commitee: Heinz, Jeffrey, Syrdal, Ann
School: University of Delaware
Department: Department of Linguistics
School Location: United States -- Delaware
Source: DAI-A 74/01(E), Dissertation Abstracts International
Source Type: DISSERTATION
Subjects: Linguistics
Keywords: Allophonic variation, English, Hidden Markov models, Parallel states, Reduced vowels, Schwa, Speech synthesis
Publication Number: 3526450
ISBN: 9781267603258
Copyright © 2019 ProQuest LLC. All rights reserved. Terms and Conditions Privacy Policy Cookie Policy
ProQuest