Early investigations into subjective audio quality assessment using brainwave responses

In this work, we take the first steps towards quantifying changes in the perceived quality of audio by directly measuring human brainwave responses using a high-resolution electroencephelograph (EEG). Specifically, human subjects are presented with audio whose quality varies with time while being monitored by a 128-channel EEG; some of the time, they move a slider bar up and down to indicate their perception of the changing quality while at other times they listen passively. Our focus here is to identify low-level features in the brainwave responses that correlate well with temporal quality variations across multiple base audio sequences and different test subjects. We consider two approaches for quality classification, one based on cross-correlation of the spectrally-averaged signals and the other based on time-space-frequency power distributions. Using a perceptually-weighted error criterion, we find that the error varies between 0.177 and 0.227 for the time-space-frequency approach and 0.0569 and 0.0723 for the cross-correlation approach. We note, however, that the latter approach is operating within the training set while the former is not, so the comparison between these two approaches is somewhat misleading.