Impact of Different Acoustic Components on EEG-Based Auditory Attention Decoding in Noisy and Reverberant Conditions

Identifying the target speaker in hearing aid applications is an essential ingredient to improve speech intelligibility. Recently, a least-squares-based method has been proposed to identify the attended speaker from single-trial EEG recordings for an acoustic scenario with two competing speakers. This least-squares-based auditory attention decoding (AAD) method aims at decoding auditory attention by reconstructing the attended speech envelope from the EEG recordings using a trained spatio-temporal filter. While the performance of this AAD method has been mainly studied for noiseless and anechoic acoustic conditions, it is important to fully understand its performance in realistic noisy and reverberant acoustic conditions. In this paper, we investigate AAD using EEG recordings for different acoustic conditions (anechoic, reverberant, noisy, and reverberant-noisy). In particular, we investigate the impact of different acoustic conditions for AAD filter training and for decoding. In addition, we investigate the influence on the decoding performance of the different acoustic components (i.e., reverberation, background noise, and interfering speaker) in the reference signals used for decoding and the training signals used for computing the filters. First, we found that for all considered acoustic conditions it is possible to decode auditory attention with a considerably large decoding performance. In particular, even when the acoustic conditions for AAD filter training and for decoding are different, the decoding performance is still comparably large. Second, when using speech signals affected by either reverberation and/or background noise there is no significant difference in decoding performance (p > 0.05) compared to when using clean speech signals as reference signals. In contrast, when using reference signals affected by the interfering speaker, the decoding performance significantly decreases. Third, the experimental results indicate that it is even feasible to use training signals affected by reverberation, background noise and/or the interfering speaker for computing the filters.