Joselyn Rodriguez / Characterizing cue and category influence in second language speech perception
Joselyn Rodriguez / Characterizing cue and category influence in second language speech perception
Thursday June 18, Joselyn Rodriguez defends her dissertation, "Characterizing cue and category influence in second language speech perception." The committee has Naomi Feldman and Bill Idsardi as Co-Chairs, Patrick Shafto (Rutgers) and alum Shannon Barrios *13 (Utah) as regular members, and Bob Slevc representing the Dean of the Graduate School. Joselyn's abstract is below.
Adults are more sensitive to the sound contrasts of their first language (L1) than second languages (L2). While researchers agree that the first language influences perception of L2 sounds, it remains a question how the L1 influences perception and whether this influence can be overcome. This thesis explores L1 influence on the L2 focusing on the role of categories and cue sensitivity on perception of an L2 using a mixture of computational modeling and behavioral experiments. The first part of the thesis addresses how L2 sounds come to be associated with categories in the L1. While L1-L2 mapping is hypothesized to be a central aspect of L2 perception, there has not yet been a successful account of how this process takes place. Addressing this, I propose a model of L1-L2 sound mapping using a mathematical framework called optimal transport (Monge, 1781; Peyre and Cuturi, 2020). I show that this model can account for two instances of L1-L2 mapping using test cases of L2 perception through a constraint over category distribution imposed by the L1. The second part of the thesis addresses the use of cues cross-linguistically. Across two behavioral experiments, I find that not all L1 cues are able to be re-used to learn a novel contrast in L2. This suggests a complex relationship between cues and the categories they identify in the L1 that is not easily dissociated when learning new contrasts. The last part of the thesis focuses on a particular hypothesis regarding how L1 specialization arises. Human infants show “attunement” towards their L1 over the first year of life, improving their perception of the L1 while losing sensitivity to non-native contrasts (Kuhl et al., 2006; Werker et al., 1981). How this attunement improves human performance is an open question, especially given the success of self-supervised speech models. Do these models show evidence of L1 attunement? In a final series of experiments, I ask whether model performance on untrained contrasts along phonetic cues displays language-specific patterns in perception similar to human listeners. Such behavior would be evidence of L1 attunement in these models. I find that while the models show specialization towards the L1, the pattern of perception isn’t comparable to human listeners, suggesting that these models learn representations that are not similar to humans.