Incremental processing of speech leads listeners to build expectations and to make predictions about upcoming sentence information. This eye-tracking study explores the time course of L2 sentence processing and the (anticipatory) integration of different types of linguistic information in different types of noise. We investigate (a) whether L2 listeners continue to predict during sentence comprehension even when noisy speech causes phonetic unreliability or difficulties in information integration, and we examine (b) the degree to which different linguistic cues are affected by different types of noise in L2 listeners. This way, we aim to investigate which cues are particularly vulnerable in L2 predictive processing and why.

Using a visual world paradigm, we tested predictive processing among 72 German advanced L2 listeners of English in three linguistic conditions (discourse, morpho-syntax, lexical semantics), and across three acoustic conditions (quiet, stationary noise, multi-talker babble noise), which were presented to three subgroups (= noise groups). Noise can impact information integration in differential ways. Whereas stationary speech-shaped noise serves as an energetic masker, lowering the salience of bottom-up information (e.g. inflectional information, function words), multi-talker babble combines energetic and informational masking, thus adding cognitive load. Language proficiency is known to modulate effects of noise on speech perception and processing, raising the question which linguistic cues are particularly vulnerable in L2 predictive processing, and how different noise types may affect the processing of these cues.

Results from cluster-based permutation analyses and lmer suggest overall effects of noise group and of linguistic condition. As expected, the use of lexical semantic information remains predictive across noise groups. Listening to speech in stationary noise leads to general delays in prediction. We also observed some interactions of noise group x linguistic condition. Comparing prediction in default vs. marked structures, our L2 listeners presented with sentences in stationary noise (compared to quiet) did not show any predictive use of inflection, suggesting processing delays based on perceptual difficulties with low-salient information. Listeners presented with sentences in multi-talker babble, by contrast, failed to predict based on discourse-related cues, suggesting that the use of discourse cues in an L2 is subject to greater difficulties when the integration of information is compromised. In all, these findings suggest that linguistic and acoustic aspects interact in predictive processing in the L2.