In one ds004703 SEEG subject (SD013), the speech envelope predicts high-gamma in 12 of 108 bipolar channels, but phonetic-feature labels at phone onsets do not add accuracy beyond shuffled labels

Does stereo-EEG (SEEG) high-gamma (HG) power during passive natural speech contain phone-identity information beyond the acoustic envelope and phone timing? We analyzed one subject (SD013, session 01) from OpenNeuro ds004703 v1.1.0: 16 passages (552 s) and 108 bipolar channels. HG was 75–115 Hz log power in 20 ms bins. Ridge temporal response function (TRF) models used 0–400 ms lags and five-fold cross-validation by passage. The acoustic model used the speech envelope and its positive derivative. A second model added 16 phonetic features at phone onsets; a control used the same onsets with phone labels permuted. The time-alignment check peaked at an extra lag of 0.00 s (r=0.10), confirming EDF time = log time + AudioOffset for this subject. Twelve channels had significant acoustic-model accuracy under a 200-shift null with Benjamini–Hochberg q<0.05; the maximum held-out r was 0.24. In these channels, true-label minus shuffled-label accuracy had median −0.0015 (4/12 positive; Wilcoxon p=0.47). True-label minus acoustic-only accuracy had median −0.0034 (p=0.034), a small loss with 336 extra regressors at this data size. Thus, in this subject and model class, about nine minutes of passive listening did not show phone-identity information beyond envelope and timing. This negative result does not establish that such information is absent. The other attempted TRF runs did not finish within the 45-minute compute limit. Limits include one subject, one HG band, a single-pass filter, no anatomical labels, phone onsets that were not checked, and lambda selection on the evaluation folds. With 200 null shifts, q values were 0.045 or higher.

stereo-EEG · high-gamma · speech · phonetic features · temporal response function · passive listening · negative result