Forward Nonlinear Model for Deep Learning of EEG Auditory Attention Detection in Cocktail Party Problem

In a multi-speaker scenario, humans are able to focus on a target speaker, ignoring all other speakers and noise, thus solving the so-called cocktail-party problem. However, elderly people and people suffering for hearing loss struggle to listening under these conditions. Recent studies have confirmed that the listener’s selective attention to the attended speaker can be decoded using recording of brain activity such as electroencephalography, thus opening new opportunities in developing a new generation of neuro-steered hearing aids and hearing prostheses. To this end several algorithms have been developed for solving the so called auditory attention decoding problem from electroencephalography on the basis of neural entrainment mechanism. The most common approaches in development of auditory attention decoding algorithms are based on linear modeling of the neural entrainment. However, even though these algorithms have shown to be effective in solving cocktail-party problem, they have some inherent limitations. The main objective of this contribution is to show that nonlinear modeling of speech-electroencephalography system ensures the best performance in terms of higher correlation between stimulus and neural response, thus proving the limitations of linear approach. For this purpose the most common linear models for auditory attention decoding are reviewed and a new nonlinear model for auditory attention decoding is proposed. An extensive experimentation using a specific speech-electroencephalography dataset, confirms the superiority of nonlinear modeling in solving the auditory attention decoding problem.