Fifteen-bin ridge readers scored 5% to 13.75% on fixed ten-word tests of trial-normalized SEEG from two people

We tested whether finer time bins improve word reading from public SEEG. We used CerebroVoice, NEMAR nm000241 version 1.0.0, from two people during overt Chinese and English speech. Each of four records had 80 fixed test trials with ten equal word classes. A linear ridge reader used three 500 ms bins or fifteen 100 ms bins per channel. Fitting rows alone set feature scales and selected the ridge penalty. With fifteen bins, high-gamma activity scored 7.5% to 13.75%. The low-frequency signal scored 5% to 11.25%. The chance reference was 10%. All descriptive paired trial intervals for the gain from finer bins included zero. The low-frequency comparison was an exploratory extension. All 2169 complete trials per signal had distinct stored signal arrays within each person. Independent byte checks and a second ridge solver passed. The package preserves compact features, fixed splits, models, every prediction, figures, code, source checksums, and retrieval instructions. The test uses complete normalized trials and reused benchmark rows. It does not test live timing, a neural cause, or stimulation. The result applies to these readers and splits. It does not show that the recordings lack word information.

stereo-electroencephalography · speech decoding · brain-computer interface · linear ridge regression · high-gamma activity · low-frequency signal · time-bin resolution · trial normalization · reproducible benchmark