Sound reduced mean test error at selected depth contacts in nine Podcast recordings with common average reference and controls for contact history and other shafts

Sound reduced mean test error by 2.88 percent after controls for target history and activity on other electrode shafts. We tested 68 depth contacts selected from the first five minutes of nine public Podcast recordings. Models used those five minutes for fitting and the last ten minutes for testing. Six participants had lower error. The 95 percent participant bootstrap interval was 1.16 to 4.54 percent. The mean gain across all 325 eligible contacts was 0.07 percent. Its interval included zero. The earlier model without target history gave a 4.60 percent mean gain and was reproduced exactly. This result applies to conditional prediction of a common-average-referenced high-gamma envelope during one story. It does not establish a neural cause or an online forecast. The package preserves compact data, source hashes, fitted models, predictions, code, figures, and retrieval instructions.

electrocorticography · speech · sound · conditional prediction · depth electrodes · high gamma · ridge regression · SEEG