EEG-based Auditory Attention Detection with Estimated Speech Sources Separated from an Ideal-binary-masking Process

Previous studies showed that auditory attention can be decoded from the corresponding electroencephalography (EEG) signals. Most existing EEG-based auditory attention detection (AAD) methods identify target speech in the competing-speaker scenes by comparing the correlation coefficients between the speech envelope of each clean stream and the speech envelope reconstructed from the EEG signals. The usage of separate speech streams limits the actualization of EEG-based AAD in the realistic environments. The current study aimed to develop and assess an EEG-based AAD method using the estimated speech sources separated from an ideal-binary-masking (IBM) process. Specially, the IBM-based speech processing method was first implemented to separate the speech sources in the competing-speaker scenes. Then the estimated IBM-processed speech sources were used to establish the AAD model and extract the target speech stream. Experimental results demonstrated that the AAD accuracies computed with the estimated IBM-processed speech sources were comparable to those with original clean speech sources over a range of signal-to-masker ratios. These findings indicate that the estimated IBM-processed speech sources provide necessary and sufficient information for the EEG-based AAD methods, which facilitate the extraction of attention-driven target speech streams in noisy environments.