Eight SEEG contacts gave a mean 0.95% error reduction beyond recent audio for 100 ms speech forecasts in ten recordings

We tested if eight SEEG contacts improve forecasts of future voice level after cue time and recent audio are available. We used public SEEG and audio from ten people with epilepsy who read Dutch words aloud. Each reader used earlier trials for fit and model choice. Later complete trials were held out. The main target was a 50 ms audio window that ended 100 ms after the latest input. The target began 50 ms after that input on corrected recorded clocks. The mean relative error reduction beyond recent audio was 0.95%. The 95% person bootstrap interval was -0.03% to 2.34%. Six of ten recordings improved. The exact person sign-flip test gave p = 0.127 under its symmetry assumption. Recent audio reduced error by 46.61% relative to cue time. The added SEEG gain was small and uncertain. The result does not establish a benefit that is consistent across people. This benchmark provides a recent-audio control for overt-speech forecasts. It does not test silent speech, a live brain-computer interface, or stimulation. The package preserves all ten feature inputs, fitted models, test predictions, scores, three figures, source hashes, and code. All ten raw recordings passed new hash checks and exact feature reconstruction. A clean rerun reproduced the numerical results exactly.

stereo-EEG · speech forecasting · voice level · audio control · ridge regression · overt speech · epilepsy · Dutch