Fixed delayed envelope regression had higher mean file error than instantaneous regression across listeners from ds004408 run 01 to run 02 under source time alignment

This study tested speech envelope prediction in OpenNeuro ds004408 version 1.0.8. Ten listeners in run 01 trained a fixed model. Nine different listeners in run 02 supplied test scores. The delayed model used envelope values from 0 to 0.5 second in the past. Both models used ridge penalty 0.1. The target was standardized 1 to 4 Hz numeric EEG on 56 fixed channels. The instantaneous model had mean squared error 1.000495890. The delayed model had error 1.000569537. The paired error decrease was -0.000073646. Its 95% person-bootstrap interval was [-0.004037566, 0.001058016]. The interval includes improvement and harm. Five of nine test people had a positive decrease. Mean channel correlation rose from 0.0002373 to 0.0103017, while mean squared error increased. All seven fixed descriptive checks are included. This adds a listener and excerpt transfer test to earlier within-excerpt covariance work. The package preserves the plan, processed inputs, code, source hashes, models, bootstrap draws, and reproduction instructions. A complete numeric repeat matched byte for byte. Four fresh source downloads matched pinned hashes and processed arrays. The result assumes source time alignment. Source physical calibration remains unresolved. It does not establish a neural delay, absence of speech tracking, or equivalent model performance.

EEG · speech envelope · cross-listener prediction · ridge regression · naturalistic speech · bootstrap uncertainty