Figure 4.

Example neural templates and utterance discrimination in Subject 1. (A) From left to right: the response spatiotemporal matched filter; an average of all keyword utterances, the bilabial spatiotemporal matched filter (STMF); an average of just keyword utterances with a bilabial place of articulation, the difference template; the subtraction of the response spatiotemporal matched filter from the bilabial spatiotemporal matched filter, and the discrimination template; the regularized and smoothed/denoised discrimination template for bilabial keywords. (B) Neural templates, created as a trial-average of particular keywords or phonemic contrasts followed by regularization and normalization, are shown for the four electrodes (a, b, c, and d) depicted in Figure 3. The VAD template, shown at the bottom, is the mean across all 120 trials in the task. The correlation of these templates with the high-gamma activity in the same task is shown in the plot to the right of the templates for a contiguous period of ~95 to ~125 s into the reading task. Vertical gray lines in this plot indicate ground truth utterance times as recorded by a microphone, and the associated utterance is indicated at the bottom of these lines. Peaks of the neural VAD output closely matched the utterance times. (C) The values of these template features across all templates (including many not pictured) at the utterance onset times were collected and reduced to two dimensions using multi-dimensional scaling, then plotted in the scatter plot, highlighting how these features clearly discriminate place of articulation and consonant voicing.