Linear ridge matched RBF decoding of the speech envelope from SEEG high gamma in six listeners

We compared linear ridge and nonlinear decoders of the speech envelope from stereo-electroencephalography (SEEG) high-gamma activity during passive listening. The preregistered plan specified 8 listeners and 12 passages per listener; 6 listeners completed decoding because the run reached its time limit. We used bipolar SEEG features (70–150 Hz Hilbert amplitude, 32 Hz), 0–250 ms backward-model lags, and held-out-passage validation. Linear ridge on all lagged channels decoded above a 200-shift chance null in each listener (mean Pearson r = 0.274; p = 0.005 each). The primary contrast, RBF kernel minus full linear ridge, was −0.009 (95% bootstrap CI −0.0229 to +0.0083; exact sign-flip p = 0.344), with higher RBF performance in 1 of 6 listeners. With matched 64-component inputs, RBF and linear ridge differed by +0.0035 (95% CI −0.0015 to +0.0099). A 64-unit MLP performed worse than full linear ridge (mean difference −0.1362; 95% CI −0.1668 to −0.1091; p = 0.031), in all 6 listeners and 69 of 72 passages. These results do not support better speech-envelope reconstruction by the tested nonlinear decoders in this dataset and setting. Limits include six listeners, about 6.7 minutes of speech per listener, offline passive perception, one RBF configuration, one small MLP, and no anatomical selection. The live-mirror dataset release is uncertain; source recordings, audio, event tables, and feature caches are not included.

speech envelope · stereo-electroencephalography · SEEG · high gamma · neural decoding · ridge regression · RBF kernel · MLP · held-out validation · passive listening