Fig 1.
Diagram of the temporal response function estimation in the forward and backward modelling approach.
The forward modelling approach or encoding model (TRF) uses a stimulus feature (i.e., speech envelope plotted in red) to predict the EEG response to a stimulus (here three TRFs and their associated EEG channels are shown). The backward modelling approach or decoding model (mTRF) uses EEG response to reconstruct the stimulus feature.
Fig 2.
(Left) The modulation spectrum of the natural speech, speech with short pauses, and speech with long pauses stimuli. (Right) The modulation spectrum of the three stimuli completely overlaps after removing both natural and inserted pauses.
Fig 3.
The mean correlation coefficient from decoders trained and tested on the full envelope in (left) delta and (right) theta bands across three speech pause conditions.
Error bars indicate the 95% confidence interval for the mean.
Fig 4.
The correlation coefficient from decoders trained on the full envelope and tested on onsets (circles) and non-onsets (squares) in (left) delta and (right) theta bands across three speech pause conditions.
Each point shows the average Pearson’s correlation coefficient across sixteen participants. Error bars indicate the 95% confidence interval for the mean. Asterisks above paired points indicate significant differences in correlation coefficients (* for p<0.01 and *** for p<0.001).
Table 1.
P-values for all possible pairwise tests (Wilcoxon signed rank tests) across all speech pause conditions and speech features tested using model trained on the full envelope for both the delta and theta bands.
Significant p-values are shown in bold and italic (critical values from Bonferroni correction). P-values which are underlined indicate that the speech feature labelled at the top of the column with an underline has significantly greater correlation coefficients, or else the other speech feature is greater.
Fig 5.
Box and Whiskers plot of correlation coefficients from each participant’s decoder using different amount of training and testing data in the delta (left) and theta band (right).
Light grey, grey, and dark grey boxes contain correlation coefficients from the full envelope using different stimulation durations (full length, 2 minutes, and 1 minute segments) recording from each data segment (in total 4 segments), respectively. Full segment stimulation refers to the full duration of the recording of each segment including natural pauses, whereas the 2 and 1 minute segments stimulation refers to recording segments with added pauses whose duration is truncated to 2 or 1 minute, respectively. The grey horizontal band indicates the range of critical values obtained from individuals in the sample, based on the null distribution of the correlation coefficients only from decoders trained and tested on the full envelope with segments in full length (i.e., all estimates above this band are deemed significant). Dots overlaid on each box are the decoder correlation coefficients from each participant. White dots indicate individual correlation coefficients that are not statistically significant based on subject’s null distribution.
Table 2.
P-values of differences in correlation coefficients between different stimulation durations (Wilcoxon signed rank test).
Bold and italic p-values indicate statistically significant difference (Bonferroni corrected) in correlation coefficients between data reduction conditions.
Fig 6.
Amplitude histogram of onset (orange) and non-onset segments (blue) in the delta (top row) and theta frequency bands (bottom row) across the four speech segments.
Fig 7.
The mean correlation coefficient from decoders trained and tested on the full envelope and envelope with pauses removed in (left) delta and (right) theta bands across three speech pause conditions.
Each point shows the average Pearson’s correlation coefficient across sixteen participants. Error bars indicate the 95% confidence interval for the mean. Asterisks above paired points indicate significant differences in correlation coefficients (*** for p<0.001).