UCLA

Individual speakers’ speech patterns differ from one another, despite presumably similar language input. What are the sources of this individual variability? In this dissertation, I explore sources of individual differences in pronunciation of coronal obstruents (/s/, /tʰ/, /cʰ/, /c/) at the ends of nouns in Korean. Research in linguistics has traditionally focused on patterns that are shared across individuals of a speech community, and speakers who deviate from the group-level generalization were treated as outliers. Research on speaker variation has primarily been in sociolinguistics, which investigates how demographic and stylistic variables affect speech. However, it has recently been recognized that individuals with homogeneous sociolinguistic background also exhibit different linguistic behavior, and there is increasing interest in how and why people vary.

One view of language learning holds that individuals who received roughly the same linguistic input as a child converge on the same linguistic knowledge (grammar), and that inter-speaker variation arises only due to random noise or extra-grammatical factors such as differences in how speakers perform the experimental task (Chomsky, 1975). In a longitudinal study involving two experiments (Chapters 4-5), I show that speaker variation cannot be dismissed as random noise or as an artifact of the task. Speakers were self-consistent in their pronunciation for the target words in the two experiments, which employed different tasks and were separated by 1.5 years. Further, the results of k-means clustering analysis in each experiment indicate that distinct speaker types indeed exist, rejecting the hypothesis that the speakers are best understood as a single group.

That individual speakers should differ from one another, despite presumably similar language input, is an interesting puzzle. In Chapter 5, I explore potential sources of the individual variability, including differences in the lexicon, grammar, cognitive traits, and pressure toward normative pronunciation. Some of these factors reliably predicted individual differences, indicating that variation is systematic.

In Chapter 6, I model individual speakers’ production patterns within the framework of Maximum Entropy Grammar (Smolensky, 1986; Goldwater & Johnson, 2003), employing Optimality-Theoretic constraints (Prince & Smolensky 1993, 2004). I provided the model with corpus frequencies as the learning data (obtained in Chapter 3) and varied the bias terms for the constraints across different learning simulations. This variability in bias terms results in different outputs from the model, replicating the individual speakers’ production of obstruent variants observed in the experiment.