Feature extraction completed for eight passive-listening SEEG subjects, but the planned spectral-versus-envelope test did not finish

This package documents a preregistered test of whether a 16-band speech spectrogram predicts intracranial SEEG high-gamma activity (70–150 Hz) better than the broadband speech envelope during passive listening to natural speech. Features were extracted for the first 12 passages from each of eight OpenNeuro ds004703 subjects (SD010, SD011, SD013, SD017, SD018, SD019, SD021, SD022), covering 90–217 bipolar channels per subject. The planned encoding analysis did not finish within a 45-minute work limit on a one-CPU host with a load average of about 27; the process received about 5–9% of that CPU. No test result is reported, and this package is not evidence for or against spectral encoding. The preregistered primary analysis compares nested ridge temporal response function models: ENV uses the mean of 16 compressed bands, while SPEC uses all 16. It specifies 0–375 ms lags at 32 Hz, leave-one-passage-out validation, alpha selection by three-fold inner cross-validation, responsive channels with ENV r ≥ 0.10, and an exact one-sided subject-level sign-flip test at α = 0.05. The plan also specifies a secondary spectral-shape decoding test, but no run of it finished. Before any result was seen, stimulus bands in the current analysis script changed from the planned fourth-order Butterworth filters, Hilbert amplitudes, and 8 Hz low-pass to an STFT of audio resampled to 16 kHz (1024-sample Hann window, 500-sample hop); the log-spaced 150–7000 Hz band edges and 0.3 power compression were retained. The decoding test was removed from the current script after the time limit stopped the first run. The plan was hashed before model fitting. Passive listening only, 12 passages per subject, and an audio-to-neural alignment based on dataset audio offsets without a separate check limit the proposed analysis. The exact dataset snapshot represented by streamed files remains uncertain. Dataset recordings, audio, event tables, and source-derived input caches are not distributed here because the saved CC0 metadata conflicts with noncommercial/no-identification terms recorded for the source collection; upstream files remain linked in the text.

SEEG · high-gamma activity · speech envelope · spectrogram · temporal response function · preregistered analysis · passive listening · OpenNeuro ds004703