Figure 3

Modeling procedure. (a) Regressors extracted from the multimedia are audio envelope and word frequency. (b) Generic modeling of continuous EEG data, including the training and testing phase. In training, we find our TRF by optimizing the error of the linear model with n-1 participants and our regressors. Then in the testing phase, predicted response is the convolution of regressors with TRF. Correlation is calculated between actual and predicted EEG responses. (c) Procedure of data analysis. (1) Both stimulus and response (EEG) are preprocessed as described in Preprocessing and Extracting of features from multimedia subsection. (2) To find the TRF, first we need to optimize our ridge regression model by finding the best lambda. With a logarithmic vector of lambda values and Eq. (2) we find TRFs for different lambda. Then by Eq. (1), we predict the response for each \(\lambda\) and then calculate the correlation between the predicted and true EEG response. Finally we choose the \(\lambda\) of the highest correlation for the training step. (3) We split data into train and test. In the individual models \(80\%\) of each subject and in Generic models n-1 subject. With the best \(\lambda\) found in the previous step and Eq. (2) we find TRF weights. (4) With the test data and founded TRF in the previous step, we calculate the predicted response with Eq. (1). Then, we calculate the correlation of true and predicted EEG response.