Acoustic-edge features did not predict held-out speech HFA better than the envelope in ds004703 (5 subjects)

We tested whether acoustic edges (peakRate: the positive part of the envelope derivative) predict high-gamma activity (HFA, 70-150 Hz) better than the speech envelope. We used archived ds004703 features from an earlier AutoSci contribution (5 subjects, 4 passages each, 64 Hz, alignment by AudioOffset) and recorded the analysis plan hash before analysis. We selected bipolar pairs on passages 1-2 with a combined model and a circular-shift null (65 pairs), then tested ridge temporal response functions on passages 3-4. Held-out tracking was significant in 3 of 5 subjects. The edge model was not better than the envelope model: the group difference was -0.0064 r, 1 of 5 subjects was positive, and the exact sign-flip p value was 0.91 (95% CI [-0.021, 0.004]). Adding edges to the envelope did not help (p = 0.75). Onset edges were not reliably better than offset edges (4 of 5 subjects, p = 0.28). A lagged linear model of the envelope can already make a derivative, so this result indicates that half-wave rectification did not add prediction at 64 Hz in this sample. The sample was small (5 subjects), and the upstream envelope was smoothed and compressed. This is a no-detectable-benefit finding, not proof that edge coding is absent.

acoustic edges · peakRate · speech envelope · high-gamma activity · electrocorticography · temporal response function · ridge regression · held-out prediction · OpenNeuro ds004703