A waveform decoder with eight SEEG pairs gave 46.3% mean balanced accuracy for heard sound classes in later English excerpts from ten people

We tested whether four signed waveform coefficients per SEEG pair improve classification of heard sounds over four band powers per pair. We used OpenNeuro ds004703 version 1.1.0, ten people, and 3,516 selected English sound events. Each method used eight pairs, a 100 to 300 ms window, and a fixed linear ridge model. We selected pairs and fit models on early excerpts. We tested on later excerpts with no excerpt ID overlap. Some name prefixes occur in both splits. Mean balanced accuracy was 46.29 percent for waveform and 47.45 percent for power. The paired difference was -1.16 percentage points. The descriptive 95 percent interval from person bootstrap draws was -6.29 to 3.82 points. The exact paired sign-flip p value was 0.6797. This test did not establish an accuracy gain. The interval does not prove equal accuracy. We checked all after-sound source hashes and rebuilt all power features. The package adds signed features before and after sound onset, source range recipes, code, model parameters, predictions, control draws, and four figures. The test concerns heard speech with known sound times. It does not test intended speech or brain stimulation.

stereoelectroencephalography · speech sound classification · waveform features · band power · balanced accuracy · brain-computer interfaces · ridge regression