In 8 ds004703 listeners, the best contact tracked decoder accuracy, while envelope tracking occurred on few shafts

We examined archived results from 8 SEEG subjects who listened to 16 speech passages. A tracking contact was defined by a maximum in-sample speech-envelope correlation r >= 0.10 with 70–150 Hz common-average-referenced activity over 17 lags from 0 to 0.5 s. Five subjects had tracking contacts; in each, they were on 2 or 3 of 9–16 shafts, and 2 shafts contained at least 80% of them. Three subjects had no contact at this threshold. Across subjects, the best contact r was associated with the held-out multichannel decoder r (Spearman rho = 0.95, exact p = 0.0011 for the high-gamma CAR decoder; rho = 0.81, p = 0.022 for the combined bipolar decoder). Median contact r was less predictive (rho = 0.67, p = 0.083; rho = 0.62, p = 0.11). In SD013, the combined bipolar decoder r = 0.30 despite a best CAR high-gamma contact r = 0.05. The sample contains 8 subjects from one dataset and one passive-listening task. Contact and decoder measures use the same recordings; contact r is in-sample and maximized over lags, while decoder r is cross-validated. The results show agreement between measures, not prediction of performance in a later session. The exact ds004703 release is uncertain. Source-derived input caches are omitted because the available saved license record conflicts with noncommercial and no-identification statements; results and figures remain available.

speech envelope · SEEG · speech decoding · high gamma · common average reference · bipolar reference · shaft concentration · passive listening · brain-computer interface