A SEEG speech detector tracked voice onset in one of two participants; cohort inference remains unresolved

We applied a pre-registered latency-slope plan to real SEEG data from two of ten participants reading Dutch words aloud. A shrinkage-LDA detector using the 70–170 Hz high-gamma envelope across contacts separated speech windows from pre-cue windows (out-of-fold AUC 0.82 and 0.89). In both participants, the voice-locked detector output crossed half of its peak about 0.15 s before audio voice onset. In sub-01, detector latency moved with reaction time (beta 0.80; 95% CI 0.36 to 1.46), consistent with production locking in this participant. In sub-02, the estimate was beta 0.23 (95% CI −0.39 to 0.93); the small reaction-time spread and wide interval leave the result unclear. The pre-registered group test gave mean beta 0.51 and p = 0.5 against both beta = 0 and beta = 1. With n = 2, the smallest possible exact sign-flip p-value is 0.5, so no group conclusion is possible. The run supports that the pipeline can be applied to real data and provides per-participant estimates; the cohort question remains open until the other eight participants are processed. Limitations include one detector over all contacts, voice onset estimated from an audio-envelope threshold without manual review, and session-wide feature z-scoring, which is not fully causal for an online BCI.

speech brain-computer interface · SEEG · high-gamma activity · speech onset detection · latency slope · reaction time · single-word production