A 16-band spectrogram did not improve prediction of SEEG high gamma over the speech envelope in passive listening

Speech BCIs often decode the sound spectrum from intracranial signals. We tested if SEEG high gamma (70-150 Hz, bipolar pairs) carries spectral speech information beyond the broadband envelope. Data: OpenNeuro ds004703 1.1.0, 8 subjects, the first 12 passages. We used the preregistered plan of sci bf8b3246 (SHA-256 7cb8fbed...). For each pair, we fit 2 nested ridge TRFs (lags 0-375 ms): ENV (mean of 16 compressed bands) and SPEC (16 bands, 150-7000 Hz). Validation was leave-one-passage-out. 43 pairs in 7 subjects were responsive (ENV r >= 0.10). SD019 had none. The mean gain r(SPEC) - r(ENV) was +0.0016 (ENV r 0.192, SPEC r 0.194). The gain was positive in 5 of 7 subjects. The exact one-sided sign-flip test gave p = 0.38, so the primary test did not show a gain. 13 of 43 responsive pairs had a significant gain by a passage test. STFT bands gave the same result (+0.0010, p = 0.41). Over all pairs, SPEC was worse than ENV (-0.004), which agrees with overfit of the larger model. Limits: passive listening only, 12 passages, linear models, high gamma only, no separate audio alignment check, 7 subjects in the group test. A null mean does not exclude spectral tuning in a few channels. The package contains the code, the provenance, the per-pair results and a figure.

SEEG · speech · high gamma · temporal response function · speech envelope · spectrogram · passive listening · OpenNeuro