A causal SEEG decoder reached 81% match accuracy at 5 s offline and peak information transfer rate at 1–2 s in 8 passive listeners

We measured how decision accuracy depends on window length for a causal ridge decoder of the speech envelope (SEEG 70-150 Hz amplitude and 1-8 Hz voltage, 250 ms latency, leave one passage out) in 8 ds004703 subjects. Task: tell the true envelope window from windows of 15 other passages. Group mean accuracy was 0.61 at 0.5 s, 0.73 at 2 s, 0.81 at 5 s and 0.90 at 20 s. Accuracy at 5 s was above 0.5 in 8/8 subjects, and accuracy at 12 s was higher than at 2 s in 8/8 subjects (exact sign-flip p = 0.0039 for each). The Wolpaw information transfer rate had its group maximum at 1-2 s windows (5.9 bits/min); two subjects with full-passage r > 0.4 reached about 17 bits/min. Subjects with higher r tended to have shorter optimal windows (Spearman rho = -0.64, one-sided p = 0.048). Exploratory: one model, p = Phi(r sqrt(9.2 W / 2)), predicted accuracy from full-passage r alone (mean absolute error 0.028 in leave-one-subject-out fits); thus the window for 80% accuracy is about 0.15 / r^2 s. 16-way passage identification from onset reached 0.58 at 20 s (chance 0.06). Scope: perceived speech, offline, distractors from other passages, one linear decoder, 8 subjects.

speech-envelope decoding · stereo-electroencephalography · decision window · brain-computer interface · information transfer rate