Fig 3. EEG reflected the linguistic transformation of speech as a function of selective-attention in a two-speaker cocktail party scenario.

Companion results split by story speaker are in Figs H and I in S1 Text. Top Row Attended: EEG prediction accuracies derived from attended speech models resembled the single audiobook data (Fig 2). Top Left Plots: Whisper complemented speech envelope-based measures and lexical surprisal in scalp-average EEG prediction. Z and FDR corrected p-values at the top of each plot correspond to signed ranks tests comparing Union EEG prediction accuracies to constituent models. Mirroring Fig 2, Whisper’s prediction accuracy increased with layer depth and latter layers complemented both the envelope measures and lexical surprisal. Top Right Scalp Maps: Electrode-wise analyses and variance partitioning (see Fig 2 caption for details) revealed that Whisper L6 dominated prediction in bilateral temporal scalp electrodes. Lexical surprisal was reflected in centroparietal electrodes. Bottom Row Unattended: EEG prediction accuracies derived from unattended speech models revealed no linguistic contribution. Bottom Left Plots: Based on scalp-average measures, Whisper added no predictive value to the envelope-based measures, and rather than increasing, prediction accuracy was weaker in later layers (see signed ranks Z and FDR corrected p-values at the top of each plot). Whisper did however improve on lexical surprisal prediction accuracies (all layers), which is presumably because it still encodes a residual of unattended speech acoustics. Bottom Right Scalp Maps: Consistent with the brain processing only superficial acoustic features of unattended speech electrode-wise analyses and predicted variance partitioning echoed that envelope-based measures alone drove prediction over central scalp electrodes.