Phrase onsets after pauses added no held-out SEEG prediction beyond the speech envelope

During passive listening to natural speech, adding a phrase-onset regressor (pause of at least 300 ms followed by at least 200 ms of speech) did not improve held-out prediction of SEEG 70–150 Hz high-gamma amplitude beyond the speech envelope in 6 of 8 subjects with envelope-tracking channels: mean Δr = +0.0007, 3 of 6 positive, exact one-sided sign-flip p = 0.23. No channel passed the family-wise shift null. An exploratory onset-locked residual was positive in 5 of 6 subjects (+0.071 z, p = 0.047, uncorrected). For speech BCIs, the results indicate no need for a separate phrase-onset term in an envelope model, while a sound-onset term after long silence can add a small gain. Post hoc comparisons showed that adding passage-start onsets produced most of that gain; under the other study’s onset definition, this pipeline gave +0.0080. The study used 8 subjects, 8 of 16 passages per subject, and offline linear models; the post hoc tests were not corrected. The source files came from an unversioned OpenNeuro S3 mirror, so the exact dataset release is uncertain. Dataset-derived input caches are omitted; results, provenance, code, and original figures are retained.

stereo-EEG · high gamma · speech envelope · phrase onset · ridge regression · passive listening · neural speech decoding