In two of eight subjects, mean decoding peaked at 4.6 minutes and was lower at 8.6 minutes; evidence is descriptive

A within-subject SEEG decoder reconstructed the speech envelope during passive listening. Across 2 decoded subjects from an 8-subject dataset, mean held-out Pearson r was 0.055, 0.060, 0.085, 0.123 and 0.119 after 1, 2, 4, 8 and 15 training passages (mean 0.60, 1.13, 2.10, 4.64 and 8.59 min). The circular-shift null was -0.002, with both subjects above it. The mean r(15)-r(4) difference was +0.034 (2/2 positive; one-sided exact sign-flip p=0.250, the minimum possible with n=2). Accuracy reached 90% of r(15) after 1.99 and 4.39 min (median 3.19 min). Adding 1-8 Hz voltage to high-gamma amplitude changed r by +0.016 at 1 passage and +0.037 at 15 passages; the difference test had p=1.000. The run budget limited decoding to 2 of 8 subjects and 4 of 16 test passages per subject, so results are descriptive and tests have low power. This study concerns passive listening and acoustic envelopes, not attempted or imagined speech, words, or phonemes. It does not establish whether accuracy continues to rise after about 8 min. The analyzed dataset release is uncertain; the archived HTTP provenance does not identify an immutable snapshot, and the dataset's saved CC0 metadata conflicts with reported noncommercial and no-identification terms.

speech envelope · SEEG · brain-computer interface · calibration · ridge regression · passive listening · learning curve