No Risky Bets: The Brain Avoids All-In Predictions During Naturalistic Multitalker Listening

Speech comprehension requires dealing with variability and uncertainty, especially when we are not familiar with the talker. Listeners achieve robust speech comprehension by tracking acoustic variability across talkers and implementing predictions based on their prior knowledge and contextual information. However, how perceptual adaptation to speech and predictive processing influence each other is still largely unknown. In this EEG study, we examined how listeners process continuous speech when exposed to a single familiar talker (Single condition) versus multiple unfamiliar talkers (Multi condition), all with native Italian pronunciation. By applying multivariate Temporal Response Function (TRF), we aimed to determine whether increased talker variability in the Multi condition influences phonemic encoding and predictive processing, as indexed by neural responses to cohort entropy (phonological uncertainty among lexical candidates) and semantic surprisal (lexical prediction error). Results showed increased neural responses to phonemic categories and to cohort entropy in the Multi than in the Single condition. These effects suggest that acoustic variability across talkers increases uncertainty in perceiving individual phonemes, prompting the brain to avoid committing to specific lexical candidates during word processing. Importantly, semantic surprisal responses were comparable between conditions, indicating that the modulation primarily affects early stages of processing. This study provided a critical test for how listeners compensate for increased bottom-up uncertainty by implementing probabilistic predictions during language comprehension. Our findings also offer insights into the flexibility of the human speech processing system in dynamically adapting to variable social and acoustic environments.