Finding Consistencies in MEG Responses to Repeated Natural Speech

Miika Koskinen

Abstract: The first steps in the attempts to unravel the perception of natural speech and to continuously follow the listener’s brain activity, are to find and characterize the perception-related phenomena and the relevant features in measured signals. In this paper, the problem was tackled by searching for consistencies in single-trial magnetoencephalography (MEG) responses to repeated 49-s audiobook passage. The canonical correlation analysis (CCA) based modeling was applied to find the maximally correlating signal projections across the single-trial responses. Using the trained model and separate test trials, projected MEG time series showed consistent fluctuations in frequencies typically below 10 Hz, with cross-trial correlations up to 0.25 (median). These statistically significant correlations between test trial projections suggest that the proposed method can extract perception-related time series from long-lasting MEG responses to natural speech.