In 10 SEEG participants, spectral shape was decodable above a shift null, while standard r reflected the envelope

We analyzed SEEG and pitch-shifted audio from 10 people who read Dutch words aloud. A linear decoder (PCA-50 plus ordinary least squares) mapped high-gamma activity (70–170 Hz) to 23-bin log-mel spectrograms. The analysis plan was written before decoding. The standard mean-bin Pearson r was almost fully explained by a predicted envelope with constant spectral shape (median M1/M0 = 0.9995), so it did not establish spectral-content decoding. A within-speech spectral-shape metric (M3) exceeded a circular-shift null in all 10 participants (median 0.17; group Wilcoxon p = 0.00098), but shape-based word identification was not above the null (median top-1 accuracy 0.015; null 0.012; p = 0.16, 0.75, and 0.90 for SYM, PAST, and FUT). The standard r was higher for neural activity at or before the sound (PAST) than after it (FUT) in all participants (p = 0.0020). For M3, PAST exceeded FUT in 8 of 10, but the preregistered comparison was not significant (p = 0.064); shape was above the shift null in 10 of 10 PAST decoders and 5 of 10 FUT decoders. Exploratory single-lag peaks were between −200 and 0 ms. Results are limited to offline linear decoding of overt speech with pitch-shifted audio; PAST activity during overt speech can include feedback from earlier frames of the same word, and these results do not establish equivalent information during imagined or attempted speech. Each participant-level null used 20 shifts, limiting the minimum p to 0.048.

speech decoding · SEEG · high-gamma activity · spectral shape · speech brain-computer interface · log-mel spectrogram · circular-shift null · Dutch word production