Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Electrophysiological evidence for prediction errors during perception of naturalistic speech
Poster Session B, Friday, October 25, 10:00 - 11:30 am, Great Hall 3 and 4
Ediz Sohoglu1, James Webb2, Connor Doyle3; 1University of Sussex
Prediction facilitates language comprehension but how are predictions combined with sensory input during perception? Previous work suggests that cortical speech representations are best explained by prediction error computations rather than the alternative ‘sharpened signal’ account. The key signature of prediction error is an interaction between bottom-up signal quality and top-down predictions. When predictions are uninformative, increasing bottom-up signal quality results in enhanced neural representations. However, the opposite occurs when predictions are informative (suppressed neural representations with increasing signal quality). Here we explore a listening situation more naturalistic than in previous work in which listeners heard sentences and predictions obtained directly from the speech signal i.e. from sentence context. In Experiment 1, listeners (N=30) heard degraded (16-channel noise-vocoded) sentences in which the context was strongly or weakly predictive of the last word, based on cloze probability. All sentences were semantically coherent. We also manipulated signal quality of the final word (2/4/8 channels). Using Temporal Response Function (TRF) analysis of EEG responses to the final word, we measured cortical representations of speech acoustic features (spectral and temporal modulations). We observed a significant interaction between context predictiveness and signal quality (F-test, p = .04). However, follow-up tests showed that there was only a marginal effect of signal quality on TRF model accuracies within the weakly predictive condition. In follow-up Experiment 2, listeners (N=31) listeners heard final words varying more strongly in signal quality (4/8/16 channels). We also included a control condition in which sentence context was unintelligible (1-channel noise-vocoded) and therefore completely uninformative about the final word. Here we observed a more robust interaction between context predictiveness and signal quality (F-test, p < .001). For the unintelligible context, increasing signal quality led to increased TRF model accuracies (F-test, p < .001) while for the strongly predictive context, increasing signal quality led to reduced model accuracies (F-test, p = .008). These findings are more consistent with the prediction error account and show that previous findings extend to more naturalistic listening situations. In ongoing work, we are extending these findings to further naturalistic materials (story listening).
Topic Areas: Speech Perception,