Figure 3 |. Changes in high-gamma activity across motor, premotor and somatosensory cortices trigger detection of speech output.

(A) Saliency analysis shows that changes in high-gamma activity predominantly from 300 to 100 ms prior to predicted speech onset (PSO) strongly influenced the nVAD model’s decision. Electrodes covering motor, premotor and somatosensory cortices show the impact of model decisions, while electrodes covering the dorsal laryngeal area only modestly added information to the prediction. Grey electrodes were either not used, bad channels or had no notable contributions. (B) Illustration of the general procedure on how relevance scores were computed. For each time step t, relevance scores were computed by backpropagation through time across all previous high-gamma frames X_t. Predictions of 0 correspond to no-speech, while 1 represents speech frames. (C) Temporal progression of mean magnitudes of the absolute relevance score in 3 selected channels that strongly contributed to PSOs. Shaded areas reflect the standard error of the mean (N=60). Units of the relevance scores are in 10⁻³.