Presentation

Craig Thorburn¹, I.M Dushyanthi Karunathilake¹, London Dixon¹, Mudi Zhang¹, Ellen Lau¹, Jonathan Z. Simon¹; ¹University Of Maryland, College Park

When hearing speech, listeners continuously anticipate upcoming concepts, words and phonemes using prior context (Ferreira & Chantavarin, 2018). Previous work has shown that expectations are formed using context at both the sublexical level (local) and sentence level (global) where representations at each level can be dissociated from one another in the neural data and are spatially separated in cortex (Brodbeck et al., 2022). We leverage the existence of such neurally dissociable representations in native speakers to investigate how naturalistic speech processing in context differs in L2 learners of English. We are recording MEG responses while native Mandarin and Korean speakers listen to an audiobook. We use temporal response function analysis (Ding & Simon, 2012) to investigate the tradeoff between a local context model where upcoming phonemes are predicted solely at the sublexical level and global context models where phonemes are predicted using sentence and lexical information. This analysis allows us to explicitly model the levels of representation of linguistic context in continuous data and investigate where these representations differ from those of native listeners. We measure English proficiency using cloze and lexical decision tasks, letting us investigate how knowledge of English impacts our results. First, we will investigate the integration of context at different levels in L2 listeners, expecting that the use of lower-level and higher-level contexts in non-native listeners will differ from native listeners due to differences in the strength of context representations. Prior literature suggests that L2 learners may not anticipate upcoming material to the extent of native listeners - particularly that they may not rely as heavily on syntactic context information to constrain predictions of upcoming material. (Hopp, 2016). This may be attributed to differing frequency distributions and the strength of lexical representations in non-native listeners (Kaan, 2014). Secondly, we test whether difficulties with non-native phonetic perception affect this integration of contextual information. Non-native speakers have difficulty distinguishing between phonetic categories that are not in their native language - particularly where two phonetic categories in the second language form only one phonetic category in the native language (Goto, 1971) - meaning these cues cannot always be utilized during sentence comprehension (Pelzl et al. 2021). Neural encoding of phonemes also scales with L2 proficiency (Di Liberto et al. 2020). This could influence anticipation of upcoming material, whereby listeners must rely more on higher-level contexts to compensate for the quality of the bottom-up signal. We predict listeners may rely more on higher-level contexts when the incoming signal is ambiguous to them, leading to a global context model more accurately predicting the neural signal in these situations. We will perform exploratory analysis aimed at teasing apart the relationship between the listeners’ native language and neural responses, taking into account phonemic inventory and phonotactic constraints. Finally, this project creates a new corpus of neural responses to naturalistic listening of speech in L2 listeners. While fMRI and EEG datasets exist, our work creates a novel corpus of naturalistic neural responses in L2 listeners that provides good spatial and temporal resolution.

Topic Areas: Multilingualism, Speech Perception