UC Berkeley

This dissertation investigates the effects of phonological neighborhoods on pronunciation variation in conversational speech. Phonological neighbors are defined as words that are different in one and only one phoneme by addition, deletion and substitution. Phonological neighborhood density refers to the number of neighbors a certain word has.

Previous research has shown that phonological neighbors impede auditory perception, but facilitate lexical production. As a result, words from dense neighborhoods are harder to perceive, but easier to produce, than words from sparse neighborhoods. Following these effects, two opposite hypotheses can be formed regarding the effects of neighborhood density on pronunciation variation. The listener-oriented hypothesis predicts that high-density words will be hyperarticulated, in order to compensate for their perceptual difficulty. But the speaker-oriented hypothesis predicts that high-density words are more likely to undergo speech reduction, as other words that are easy to produce (such as high-frequency words).

To test these hypotheses, two statistical models are constructed to investigate neighborhood effects on pronunciation variation in CVC monomorphemic content words. All speech data come from the Buckeye Corpus of Conversational Speech (Pitt, Dilley, Johnson, Kiesling, Raymond, Hume, and Fosler-Lussier 2007), which contains audio recordings of 40 speaker's free-form interviews. The dataset for the current research consists of more than 500 target words, represented by over 13,000 tokens. The outcome variables in the two models are word duration and degree of vowel dispersion, respectively. Neighborhood density and average neighbor frequency are the critical predictors in both models. Other factors that might affect pronunciation variation in conversational speech, such as word frequency and predictability, are entered into the models as control factors.

The results of the model analyses show that everything else being equal, high-density words are realized with shorter durations and more centralized vowels than low-density words. These findings provide strong evidence for the speaker-oriented hypothesis. A peripheral finding in the current research is a tendency for words with high-frequency neighbors to have more dispersed vowels. However, this effect is only significant when neighbor frequency from the Hoosier mental lexicon (Nusbaum, Pisoni, and Davis 1984) is used, but not when neighbor frequency computed from the CELEX database (Baayen, Piepenbrock, and Rijn 1993) is used.

Overall, major findings of the current research support the speaker-oriented hypothesis over the listener-oriented hypothesis, which suggests that word-level pronunciation variation is more heavily influenced by features of the speaker's own production system than by the consideration of listeners' needs.