Added speech trials did not show a clear gain for a fixed imagined word decoder in one SEEG participant

We tested whether added spoken and mimed trials help after three imagined training trials per word. Public VocalMind SEEG data supplied 342 trials from one participant and 19 complete word items. Each test excluded one repetition group from every training mode. A fixed ridge model used 1,980 log power features per trial. Imagined-only training gave 7.37 percent mean test accuracy. Added spoken and mimed trials gave 7.98 percent. The paired change was 0.61 percentage points. A 95 percent word-item bootstrap interval was -2.98 to 4.12 points. Correct source word labels did not clearly beat random source labels (P = 0.1322). A spoken control gave 17.54 percent accuracy and passed its label test (Holm P = 0.002). Added spoken trials alone gave 10.61 percent imagined accuracy, with an uncertain paired gain. The package preserves features, raw sample checks, 870 fits, label controls, figures, code, and source records. The evidence is limited to this participant, these visually cued word items, and this fixed model.

SEEG · speech decoding · brain-computer interface · calibration · ridge regression · VocalMind