Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Adaptive Prediction: The Brain Trades Phonemic for Semantic Expectations Under Increased Acoustic Variability
Poster Session D, Thursday, October 1, 4:30 - 6:30 pm, Wangari Maathai
Giorgio Piazza1, Marco Sala3, Rebecca Guerrini2, Martin Winchester1, Francesca Peressotti2; 1Trinity College Dublin, 2University of Padova, 3University of Geneva
Humans interact daily with different talkers, many of whom are unfamiliar and vary widely in how they produce speech. For example, during a conference question time, listeners may be exposed to a rapid succession of unfamiliar talkers that differ in accent, voice quality, fundamental frequency, speech rate, and articulation style. Despite this variability, listeners typically comprehend speech with remarkable ease. One explanation for this efficiency is that comprehension involves not only the bottom-up integration of incoming information, but also top-down predictions about upcoming input that are guided by contextual information and internal knowledge. These predictions are thought to operate across multiple levels of representation, from phonemes to words and sentences. However, much of the evidence supporting predictive processing comes from studies that minimize acoustic variability, often relying on a single talker and highly controlled speech. Therefore, it remains unclear how acoustic variability influences the brain’s ability to generate and use predictions during naturalistic speech comprehension. In this study, electroencephalography (EEG) signals were recorded from thirty Italian native participants as they listened to continuous stories under two conditions. In the Single-talker condition (hence Single) one speaker narrated the entire story; in the Multi-talker condition (hence Multi), different speakers narrated different sections of the story, introducing substantial acoustic variability and requiring participants to adapt to changes in speaker’s voice. Temporal Response Function analysis was used to map the relationship between stimulus features and neural responses over time. We analysed the neural encoding of phonemes, phonemic surprisal (i.e., the unpredictability of the next phoneme), and semantic surprisal (i.e., the unpredictability of lexical elements) over time as these measures allow us to test how variability at the acoustic-phonetic level interacts with predictive processing across the linguistic hierarchy. We observed stronger neural responses to phonemes and reduced responses to phonemic surprisal in the Multi than in the Single condition, indicating greater speech perception demands and weaker phonemic predictions. Conversely, semantic surprisal responses were stronger in the Multi condition, suggesting increased reliance on lexical-semantic predictions. Together, these results reveal a trade-off in predictive processing across levels of representation: increased acoustic variability reduces the strength of phonemic predictions while enhancing reliance on semantic expectations. This pattern suggests that predictive processing is dynamically weighted rather than globally disrupted: when bottom-up information becomes less reliable, the system reallocates resources toward higher-level sources of information that can better constrain the mapping of the acoustic signal into linguistic categories. To conclude, our findings highlight the adaptive nature of predictive processing in natural speech, which allows listeners to maintain robust comprehension in complex and variable listening environments.
Topic Areas: Speech Perception,