Five SEEG feature times reduced mean squared log speech amplitude error versus one selected time on later passages in ten recordings with calibrated penalties

We compared two readers of heard speech amplitude in ten public SEEG recordings. The source was OpenNeuro ds004703 version 1.1.0. Both readers used the same eight adjacent contact pairs. One reader combined five feature times. The other used one time selected during calibration. Each reader had its own calibrated ridge penalty. Training used earlier passages. Testing used later passages. The target was log root mean square sound amplitude with a 400 ms readout delay. Five times reduced mean person test error by 6.70 percent. The descriptive 95 percent person bootstrap interval was 2.60 to 10.82 percent. Eight of ten people improved. The exact two-sided sign-flip p value was 0.015625. This test assumes independent people and symmetric null changes. The mean gain reversed at five of six shared penalty settings. Both readers beat a constant prediction in only five people. Thus, the gain applies to the tested calibration methods. It does not isolate a general benefit from temporal inputs. This exploratory study reuses 95 clinical passage segments. It does not establish intended speech readout or live brain control. Fresh public retrieval matched all 229 data files exactly. The SCI package includes authored code, results, fitted readers, predictions, and three figures. It omits dataset recordings, copied metadata, and source-derived input caches because the source README restricts distribution; these inputs are linked at the versioned source URL. Public replay from stored inputs is therefore limited.

stereoelectroencephalography · speech amplitude · ridge regression · temporal features · neural decoding · cross-validation