In eight passive-listening SEEG participants, causal speech-envelope decoding gains accuracy from 0 to 250 ms latency

We measured held-out speech-envelope decoding accuracy across output latencies using causal SEEG features in eight participants listening passively to the first 10 passages in the order recorded for OpenNeuro ds004703. The mean Pearson r for the combined causal high-gamma and 1–8 Hz features rose from 0.163 at 0 ms to 0.225 at 250 ms (difference 0.062; 7/8 participants positive; exact one-sided sign-flip p = 0.012). Causal high-gamma-only accuracy was nearly flat after 62.5 ms (r = 0.192–0.199); 1–8 Hz-only accuracy was near zero through 125 ms and reached 0.135 at 469 ms. At 469 ms, the acausal-minus-causal difference was 0.012 (5/8 positive, p = 0.17), so the planned test did not support a reliable accuracy cost from causal filtering at that latency. The median latency to reach 90% of each participant’s 469 ms accuracy was 156 ms; the three highest-accuracy participants reached this level at 0–62.5 ms. Latencies are relative to dataset audio alignment, which may have a constant error of about 0.1 s. Accuracy at latencies at or below 0 ms partly reflects speech-envelope autocorrelation. The study used passive listening, 10 passages per participant, ridge models with 16 PCA components per feature type, and passage-wise Pearson r; it was not a streaming test with fixed calibration. The source report records dataset README restrictions on commercial machine-learning use and re-identification, while cached metadata for two candidate versions lists CC0; the exact release used is uncertain.

SEEG · speech envelope · causal decoding · output latency · high gamma · electrocorticography · ridge regression · passive listening