In six SEEG participants, a causal high-gamma speech detector followed voice onset more than the word cue

We tested whether a causal 70–170 Hz SEEG speech detector followed voice onset or a visual word cue in six participants reading Dutch words aloud. The primary latency slope (β) was the change in detector t50 divided by the change in reaction time; β = 0 represents cue locking and β = 1 represents voice locking. All six participants had β > 0. The mean β was 0.73, and the exact two-sided sign-flip test rejected β = 0 (p = 0.031, the smallest possible p at n = 6) but did not reject β = 1 (p = 0.16). Median detector AUC was 0.90. A log-space common-average reference sensitivity analysis gave mean β = 0.67 and the same p-values. These data are consistent with detector output that mostly follows voice onset, with a possible small cue-locked component; this sample cannot distinguish that pattern from voice locking. The median voice-locked t50 was −0.061 s in the four new archived-feature runs. The group combines two pipelines: raw data with CAR for two participants and archived 10 ms features without CAR for four. The archived runs used a mel-power onset rule and a NumPy LDA implementation that was not compared with scikit-learn. Four of ten dataset participants were not tested. The short lead time means the detector may gate self-paced speech but does not support prediction long before voice onset.

speech brain-computer interface · SEEG · high-gamma · voice onset · latency slope · speech detection