SEEG activity decoded a small part of speech spectral shape beyond the linear envelope in six listeners

During passive listening to natural speech, common-average-referenced SEEG high-gamma amplitude (70–150 Hz) and 1–8 Hz voltage decoded a small part of speech spectral shape beyond a linear model of the broadband envelope and mel mean. In six subjects with 16 passages each, leave-one-passage-out backward ridge models decoded the residual 16-band spectral shape at mean Pearson r = 0.031, versus 0.002 for a mismatched-passage null. All six subjects were above null, with Wilcoxon p < 0.05 in all six; an exact one-sided sign-flip test gave p = 0.016, the minimum possible with six subjects. The effect was about 13% of envelope decoding (mean r = 0.232; null 0.009). Residual decoding was largest in the lowest and highest mel bands and near zero at band 8 (p = 0.69), consistent with a low-versus-high spectral balance. Residual PC1 decoded at r = 0.056; PC2 at 0.032; PC3 at 0.011 and did not meet the group sign-flip threshold (p = 0.17). The study used six of eight planned subjects because extraction for SD019 and SD022 stopped after 7 of 16 passages. Listening was passive, models were linear, residualization removed only linear envelope effects, and zero-phase filters were non-causal. The mismatched-passage null preserved autocorrelation but not passage-specific spectral statistics. This is an encoding result, not a real-time decoder test.

stereo-electroencephalography · speech · spectral shape · high gamma · backward ridge decoding · passive listening