Cortical tracking of speech reveals top-down reconstructive processes

Cortical entrainment to the (quasi-) rhythmic components of speech seems to play an important role in speech comprehension. It has been suggested that neural entrainment may reflect top-down temporal predictions of sensory signals. Key properties of a predictive model are its anticipatory nature and its ability to reconstruct missing information. Here we put both these two properties to experimental test. We acoustically presented sentences and measured cortical entrainment to both acoustic speech envelope and lips kinematics acquired from the speaker but not visible to the participants. We then analyzed speech-brain and lips-brain coherence at multiple negative and positive lags. Besides the well-known cortical entrainment to the acoustic speech envelope, we found significant entrainment in the delta range to the (latent) lips kinematics. Most interestingly, the two entrainment phenomena were temporally dissociated. While entrainment to the acoustic speech peaked around +0.3 s lag (i.e., when EEG followed speech by 0.3 s), entrainment to the lips was significantly anticipated and peaked around 0-0.1 s lag (i.e., when EEG was virtually synchronous to the putative lips movement). Our results demonstrate that neural entrainment during speech listening involves the anticipatory reconstruction of missing information related to lips movement production, indicating its fundamentally predictive nature and thus supporting analysis by synthesis models.