Conference on Cognitive Computational Neuroscience

A growing body of research is concerned with modelling magneto- or electroencephalography (MEEG) responses to speech. This usually focuses on predicting the low-frequency time-domain portion of the signal in the delta (1-4 Hz) and theta (4-8 Hz) ranges from different stimulus features. Here, we re-analyse an MEG dataset of passive story listening and instead focus on oscillatory power. We obtain spatio-spectral response filters from a canonical correlation analysis (CCA) that maps MEG sensor power time courses onto the time-lagged envelope of the speech stimulus. These weights indicate generators in the beta frequency range (13-30 Hz) in bilateral auditory cortices, which we confirm with source localisation. We then predict the projections of the MEG responses through the spatio-spectral filters with features of a recurrent network that predicts distributions of the future of the speech envelope. We find that network outputs related to the variance but also the mean of the upcoming sound energy predict the MEG response component better than acoustic baseline models.