Fig. 5:

Quantifying and mitigating acoustic contamination of neural signals.

(A) Spectrograms for audio and neural data in the electrode and block exhibiting the strongest audio-neural correlations. Frequencies range from 5 to 1000 Hz. The bottom plot shows the same electrode after LRR “decontamination”. (B) Plot of the mean audio PSD (red) and all electrodes’ Pearson correlations (blue) from the same example block. Inset shows correlation coefficients of individual electrodes (rows) across frequencies (columns). Black horizontal ticks denote electrodes excluded from neural analyses. The pink arrow shows the example electrode from panel A. (C) Change in audio-neural correlations after LRR, pooled across all blocks, electrodes, and frequencies (restricted to electrodes with r² > 0.1 originally). Values to the right of the dotted ‘0’ line indicate a reduction in correlation strength. The mean audio-neural correlation reduction was 0.26. (D) Full classifier confusion matrix after LRR (25.8% overall accuracy across 39 classes). (E) Confusion matrix for first phoneme decoding after applying LRR. As in D, the classifier used a 500 ms window centered on voice onset. (F) Confusion matrix showing decoding each word’s first phoneme using 500 ms leading up to voice onset to avoid possible audio contamination or neural activity related to auditory feedback.