Fixed speech models had mean squared test error of 0.919 with past and current amplitude and 0.958 with future and current amplitude in five SEEG patients

We tested whether past speech amplitude gave lower error than future amplitude in models that also included current amplitude. We used public SEEG from OpenNeuro ds004703, version 1.1.0. One contact pair per patient was selected with early passages only. Fixed ridge models were fitted on 20 early passages and tested on 23 later passages from five patients. Mean squared error was 0.91944 with past and current amplitude and 0.95810 with future and current amplitude. Four patient differences favored past amplitude. The mean difference was 0.03866. Its 95% t interval was -0.00115 to 0.07847, and the exact two-sided sign flip p value was 0.125. A 40 ms assumed audio shift reduced the mean difference to 0.01360. The group result was uncertain. This was a new analysis of previously used recordings. It did not establish a neural cause. All 43 selected passage features were rebuilt exactly from pinned public source bytes during the original source check. Independent calculations reproduced the predictions and errors. The SCI package retains the text, figures, numerical results, provenance, and compatible code. Source-derived input arrays and copied source metadata are omitted because the dataset metadata and README state conflicting reuse terms; reproducing the analysis requires obtaining the source inputs under their terms.

stereoelectroencephalography · speech amplitude · high gamma power · ridge regression · temporal prediction