A rank three SEEG reader used 6016 coefficients and gave 11% to 12% word accuracy under repetition splits in one participant

We tested a fixed rank three reader on public VocalMind SEEG data from one participant. Each task used 19 complete word items with six repetitions. Rank three was fixed before new model scores were inspected. Under repetition splits, it used 6016 coefficients in a 19-score format and classified 13/114 spoken, 13/114 mimed, and 14/114 imagined trials correctly. Full ridge used 37639 coefficients and classified 20/114, 23/114, and 12/114 trials correctly. Rank three passed three word-label association tests after Holm correction under within-repetition label shuffles. Its accuracy gains over time averaging had conditional word bootstrap intervals that included zero. A later contrast-storage audit found rank three used slightly more coefficients than time averaging in that format. The checked streaming implementation retained three latent sums between windows and matched all 342 primary-trial decisions. These data do not establish performance in new participants, broad imagined-vocabulary decoding, or an online device. Shared visible word cues and uncertain recording chronology limit interpretation.

stereoelectroencephalography · speech decoding · reduced rank regression · word classification · brain-computer interface · repetition split · model storage · streaming inference