Speech envelope tracking under documented AudioOffset alignment in four of five ds004703 subjects

We tested whether bipolar high-frequency activity (HFA, 70–150 Hz) tracked the acoustic envelope of natural English speech in OpenNeuro ds004703, using the documented alignment EDF time = log time + AudioOffset. The plan was written and hashed before statistics were computed. Because downloads were slow, we analyzed the first four English Block 1 passages in five subjects (about 2.5 minutes per subject). For each subject, the statistic was the maximum across bipolar pairs of the Pearson correlation over lags of 62–250 ms; significance was assessed against 500 circular shifts of each envelope. Four of five subjects had p < 0.05 (binomial p = 3.0 × 10⁻⁵). After selecting a pair on passages 1–2, held-out correlations on passages 3–4 were 0.393, 0.158, 0.021, 0.026 and 0.287 (mean 0.177; one-sided exact sign-flip p = 0.031, the smallest possible with five subjects). In SD012 ses-02 and SD022, the held-out lag peaks were +141 ms and +125 ms, respectively. These results support envelope tracking in this limited sample and are inconsistent with an alignment error greater than 2 s for SD012 ses-02, SD015 and SD022. They do not establish that alignment accuracy for the full dataset is about 100 ms. Only five subjects and four passages per subject were analyzed; contacts were not localized anatomically, SD012 contacts are ECoG, and the circular-shift null does not control for a common passage-onset response. The binomial test assumes independent subject permutation p-values, and the held-out test has only five subjects.

speech envelope · high-frequency activity · SEEG · ECoG · AudioOffset · electrophysiology · natural speech · OpenNeuro