Characterize degradation of envelope-reconstruction AAD methods under rigorous cross-validation

Determine the extent to which EEG-based speech-envelope reconstruction decoders—specifically Wiener Filter and Canonical Correlation Analysis models—experience performance degradation when evaluated under rigorous cross-validation schemes such as leave-one-trial-out and leave-one-subject-out, in order to quantify their susceptibility to trial fingerprints and chronological biases in auditory attention decoding tasks.

Background

Auditory attention decoding (AAD) studies have shown that many models overfit to trial-specific fingerprints when evaluated with within-trial splits, prompting the use of more rigorous cross-validation schemes like leave-one-trial-out or leave-one-subject-out to obtain unbiased accuracy estimates. Prior work has documented substantial performance degradation for classification-based AAD models under these rigorous protocols, highlighting vulnerability to chronological and trial-specific biases.

However, whether and to what extent envelope reconstruction approaches—such as Wiener Filter and Canonical Correlation Analysis, which predict the attended speech envelope from EEG—exhibit similar degradation was identified as not yet characterized at the time of writing. The paper introduces a cEEGrid dataset and experiments designed to probe attentional switching and rigorously validate models, motivating a formal characterization of degradation for envelope reconstruction methods under leave-one-out-style evaluations.