Contact mean removal retained spoken and mimed word label association for a fixed classifier in one SEEG participant under exchangeable labels

This study tests contact mean removal for a fixed word classifier in one public SEEG participant. It uses 342 trials, 19 word items, and 110 contacts from VocalMind. The classifier trains on five filename repetition groups and tests on the sixth. Mean removal gave 14.04 percent accuracy for spoken trials and 14.04 percent for mimed trials. Chance was 5.26 percent. Both passed label controls with Holm P = 0.0105 across nine primary tests. Full features gave 17.54 and 20.18 percent. The test assumes exchangeable labels within repetition groups. Recording order is not established. Shared visual cues and movement can affect the result. No imagined primary score passed its control. A proof shows invariance to common log-power offsets and nonzero trial voltage gain. Numerical checks confirmed this property. All 357 source trial features were rebuilt from pinned public raw files and matched exactly. The package preserves code, feature data, source hashes, all model scores, label controls, four figures, and reproduction instructions. The result supports a gain-invariant readout benchmark. It does not establish a useful speech interface or a speech-specific neural source.

stereoelectroencephalography · brain-computer interface · word classification · log power · contact mean removal · gain invariance · permutation test · spoken speech · mimed speech · VocalMind