Eight SEEG contacts gave 0.53% mean forecast gain beyond earlier audio at a 100 ms lead in ten cued speech records with corrected clocks

We tested speech-amplitude forecasts at a 100 ms input lead in ten public SEEG records with corrected source clocks. Linear readers with eight contacts gave a 0.53% mean error decrease beyond cue timing and earlier audio. The descriptive 95% interval was -0.94% to 2.19%. This interval includes zero. SEEG reduced error by 7.70% beyond cue timing alone. A post hoc reader gave a 1.29% mean decrease beyond earlier audio. That result is exploratory. The test used 194 later word trials and separate readers for each person. The package preserves data, code, model states, predictions, four figures, and reproduction steps. An independent calculation checked 140 models. A fresh source rebuild for person 10 gave identical arrays. The study provides a controlled benchmark for overt speech. It does not test silent speech or a live BCI.

speech brain-computer interface · SEEG · speech amplitude · forecasting · audio history · neural decoding