Speech-envelope decoding improved with each added training passage in 8 passive-listening SEEG subjects

We measured how held-out speech-envelope decoding accuracy changed with calibration speech in 8 subjects listening passively to 8 passages each. A backward linear decoder used bipolar 70–150 Hz SEEG amplitude, 32 channel principal components, 6 neural lags (0–469 ms), and ridge regression with an inner cross-validation choice of alpha. Training on 1 to 7 passages (0.57 to 4.04 min) increased group mean held-out Pearson r from 0.151 to 0.242. The gain from 1 to 7 passages was +0.091 (8/8 subjects, exact one-sided Wilcoxon p=0.0039); gains remained at 7 versus 4 passages (+0.024, 7/8, p=0.0078) and 7 versus 6 (+0.006, 8/8, p=0.0039). The slope was positive in 8/8 subjects (+0.032 r per doubling, p=0.0039). The curve reached 90% of its value at 4.04 min after 2.28 min (bootstrap 95% CI 1.92–2.93 min), but continued to rise at the last step, so this relative value does not show a plateau. Two subjects remained near chance; 6 of 8 had a usable decoder. The study covers passive listening, one session per subject, bipolar high gamma, one linear decoder family, and at most 4 min of training speech. The exact ds004703 release used by the archived features is uncertain.

speech envelope · SEEG · speech decoding · high gamma · calibration · learning curve · ridge regression · neural decoding