Four time bins gave no gain over repeated mean log power on 40 later sentence trials in one SEEG record

We tested whether four time bins improved identification of ten known sentence items. We used W2 from public SEEG dataset OpenNeuro ds004993, version 1.1.2. The first 60 trials fitted readers. The last 40 trials tested them. Each speech reader used eight adjacent contact pairs and one-second windows. Both readers had 32 input columns. The mean reader repeated each mean log power four times. Pair selection and scaling used only fitting trials. Each speech reader got 9 of 40 later trials correct, or 22.5%. The paired gain was zero percentage points. A 95% block-resampling range was -10 to +10 points. A separate reader before sound got 6 of 40 correct. These results do not establish equal future performance. The task supplies onsets and ten known candidates. Items can differ in speaker, sound, and content. The package preserves features, source versions and hashes, sample audits, code, models, predictions, three figures, and reproduction instructions. Independent calculations and a full source replay passed. This test concerns listening in one person. It does not test intended speech or direct brain stimulation.

stereoelectroencephalography · sentence identification · speech · brain-computer interface · ridge regression · time-resolved power