Two Hz updates reduced mean sound amplitude error skill compared with 10 Hz in ten offline SEEG readers with eight pairs, separate phase scores, and 400 ms delay

This study tested slower feature delivery in ten people who heard speech. It used public SEEG data from OpenNeuro ds004703 version 1.1.0. Each reader used eight contact pairs and a 400 ms sound-to-readout delay. The study fitted readers on 57 earlier passages and scored 38 later passages, with 4218 unique test targets. Reducing updates from 10 Hz to 2 Hz reduced the nominal feature payload by 80%. The slower reader held the last values and was fitted again. Each clock phase was scored separately. Mean error skill fell from 5.35% to 1.60%, a change of -3.75 percentage points. Seven people had a loss. The person bootstrap interval was -7.76 to -0.17 points. The exact sign flip reference gave p = 0.119. The result shows a mean loss in this cohort; it does not establish a population loss at the 0.05 sign flip level. Fitting for 2 Hz improved mean skill by 13.32 points over using the fixed 10 Hz fit at 2 Hz. Nine people had this gain. The package includes source code, derived inputs, range records, all scores, fitted models, predictions, uncertainty samples, repeat instructions, and four figures. Independent source math and reader checks passed. Fresh source checks rebuilt the first selected passage of every person. This is an exploratory offline heard-speech study. It gives no evidence about device energy, live intended speech, or brain stimulation.

stereoelectroencephalography · brain-computer interfaces · speech amplitude · feature update rate · ridge regression · offline decoding · signal processing