In six Dutch SEEG participants, a high-gamma detector tracked speech timing more than cue timing

We applied the pre-registered latency-slope test to SEEG recordings from 6 of 10 participants reading Dutch words aloud. A shrinkage-LDA detector used 70–170 Hz log high-gamma features from all contacts and a causal 100 ms window. We split trials at the median reaction time and measured the shift in detector half-rise time relative to the shift in reaction time. All 6 detectors had AUC 0.82 to 1.00 and were retained. All 6 latency slopes were above 0 (range 0.23 to 1.19), with 4 of 6 bootstrap confidence intervals excluding 0. Mean slope was 0.73. The exact sign-flip test rejected a pure cue-locked detector (p = 0.031, the smallest possible value for n = 6), but did not reject a pure production-locked detector (p = 0.094). Detector output crossed half of its peak 0.02 to 0.15 s before audio-threshold voice onset. These interim results indicate that detector timing follows speech timing in this cohort and support further study of self-paced speech-onset detection in a BCI. They do not exclude cue-related or planning activity. Limitations include an interim sample of 6, small reaction-time spread, one detector across all contacts, audio-threshold voice onset, whole-session z-scoring, and overt speech only.

SEEG · speech production · speech onset detection · brain-computer interface · high gamma · latency slope · Dutch · pre-registered analysis