Full prefix offset correction increased mean sound amplitude error in ten fixed offline SEEG readers

We tested whether an early passage prefix improved sound amplitude estimates from public SEEG data. Ten people had separate fixed linear readers with eight pairs and a 400 ms readout delay. The prefix used ten active speech feature frames. It did not use test sound targets. Full correction increased equal-person mean squared error from 1.4831 to 1.6756 on 38 later passages. Mean error skill decreased by 11.82 percentage points. Eight of ten people had lower skill. The mean loss remained 10.74 points after digital-limit passages were removed. A fraction selected on one earlier validation passage reduced the observed harm but gave no clear gain. The correction changed passage bias only. Centered correlation did not change. A proof and numerical checks show exact protection against constant additive feature offsets that affect both the prefix and later input. That property did not improve error on these unmodified passages. The public package preserves fitted models, predictions, checks, code, and four figures. Source recordings, audio, event tables, source metadata, native sample extracts, and cached extracted features are omitted because the dataset metadata and README state conflicting terms; source URLs, hashes, methods, and reconstruction steps are retained. The analysis reuses data and some sounds are shared; source-derived inputs are omitted from the public package and must be obtained from OpenNeuro under its stated terms. The uncertainty values are descriptive. The result applies to offline heard sound and this calibration rule.

stereoelectroencephalography · brain-computer interface · sound amplitude · calibration · domain adaptation · ridge regression · feature offset · error skill