Four voltage means with eight SEEG pairs gave 49.8% mean balanced accuracy for heard sound classes in later English passages from ten people

We tested a small readout for heard vowels and unvoiced fricatives in public SEEG data from ten people. We used OpenNeuro ds004703 version 1.1.0. Each person had a separate model. The model used four signed voltage means from each of eight pairs in a 200 ms window. Pair selection and fitting used early passages. The test used 1,484 events in later passages. Mean balanced accuracy was 49.8% for voltage and 47.4% for four power bands. The paired mean change was 2.3 percentage points. Its 95% person bootstrap interval was -1.6 to 6.4 points. The exact two-sided sign-flip p value was 0.326. The mean voltage result did not exceed its planned label control (p = 0.566). This test does not show a reliable gain from the voltage method. It does not prove equal performance or absence of speech information. The package preserves features, raw-range retrieval code, source versions, hashes, fitted models, test scores, controls, and three figures. All prior power features were rebuilt exactly. An independent EDF reader checked both sample rates. The complete numerical analysis was repeated. The result concerns heard sound classes. It does not test intended speech, live control, or stimulation.

SEEG · speech perception · vowels · fricatives · brain-computer interfaces · balanced accuracy · voltage features · signal processing