A Multi-subject Temporal-spatial Hyper-alignment Method for EEG-based Neural Entrainment to Speech

The low signal-to-noise ratio (SNR) of a neural recording is typically improved by averaging the neural response over repeated trials. However, it is not applicable when studying neural entrainment to speech stimuli, in which stimuli are presented only once. Alternatively, multiple subjects' neural responses to the same stimuli can be averaged to decrease unexpected noises caused by breathing, lack of attentiveness etc., excluding those caused by heartbeats and blinking during long auditory tasks. However, individual differences such as varying latency times of the neural response to the stimulus and electrode positioning in the setup reduce the effectiveness of this method. To eliminate individual differences, we first estimated the importance (weight) of each electrode using a spatial filter and maximized the SNR by adjusting the latency of the neural response. Then we extracted the common response for all subjects and constructed neural entrainment models accordingly. The correlations between the predicted and actual neural responses obtained in this study were much higher than that in other methods in the forward neural encoding process. In the decoding process, the correlation between the reconstructed speech envelope and the original also increased significantly in both the delta and theta bands compared with previous studies.