Mixed music and rest calibration gave an uncertain mean background decrease in nine fixed SEEG speech readers

We tested whether early rest data improved thresholds for nine fixed SEEG speech readers. We used public OpenNeuro ds003688 version 1.0.7 data. Music calibration used 104 windows. Mixed calibration used 52 music and 52 early rest windows. Each window lasted 0.5 s. Separate later music, rest, and heard speech windows supplied the tests. The mean of the larger patient background rates changed from 8.59% to 7.80%. The paired change was -0.79 percentage points. Its 95% patient bootstrap interval was -7.00 to 3.43 points. Six of nine patients had an increase. All planned sensitivity check intervals crossed zero. A rule that used the larger of two separate background thresholds reduced mean background outputs to 2.91%. It also reduced heard speech sensitivity from 8.87% to 4.59%. The test did not establish a mean benefit from mixed calibration. The package supplies a fixed threshold benchmark, three figures, code, derived features, raw audit windows, source hashes, frozen models, and every test decision. Independent signal and score checks passed. Fresh full film and rest retrieval for one patient gave exact array agreement. A separate process reproduced eight numeric result files exactly. This extension reuses source patients and models. It tests passive heard speech and short rest records. It adds no independent patient replication and does not test intended commands or brain stimulation.

SEEG · speech detection · brain-computer interface · threshold calibration · background activity · rest · music · signal processing · epilepsy