Spatial streams selected from early data reduced the mean transmitted value count by 93.1 percent in nine fixed offline 400 ms SEEG readers

We tested spatial streams for fixed SEEG readers of heard sound amplitude using public OpenNeuro ds004703, version 1.1.0. Nine people supplied 54 early training passages and 100 later test records. Each full reader used 24 current features from eight bipolar pairs and had a 400 ms output delay. Early data selected one to five spatial streams, with an equal-person mean of 1.67. This reduced the mean transmitted value count by 93.1 percent at equal precision, before packet overhead. Mean squared error rose by 0.60 percent on the 100 later records; the descriptive 95 percent person bootstrap interval was -0.26 to 1.66 percent. The largest person-level rise was 3.96 percent. Mean within-passage correlation changed from 0.343 to 0.338. Five streams reproduced each full reader within numerical precision. These data had been studied before; this is a new compression test, not a new clinical sample. All eight pairs still require acquisition. The test retains the temporal filter at the receiver and does not measure implant power, intended speech, live control, or brain stimulation. Result tables, predictions, figures, proof, checks, and MIT code are included. Source recordings, metadata, audio, and derived input caches are omitted because the dataset's CC0 metadata conflicts with its noncommercial and no-identification README restrictions. The publication does not resolve those source terms.

SEEG · spatial compression · signal processing · brain-computer interface · sound amplitude