An LMMSE-based Estimation of Temporal Response Function in Auditory Attention Decoding.

Abstract

One of the remarkable abilities of humans is to focus the attention on a certain speaker in a multi-speaker environment that is known as the cocktail party effect. How the human brain solves this non-trivial task is a challenge that the scientific community has not yet found answers to. In recent years, progress has been made thanks to the development of system identification method based on least-squares (LS) that maps the variations between the cortical signals of a listener and the speech signals present in an auditory scene. Results from numerous two-speaker experiments simulating the cocktail party effect have shown that the auditory attention could be inferred from electroencephalography (EEG) using the LS method. It has been suggested that these methods have the potential to be integrated into hearing aids for algorithmic control. However, a major challenge remains using LS methods such that a large number of scalp EEG electrodes are required in order to get a reliable estimate of the attention. Here we present a new system identification method based on linear minimum mean squared error (LMMSE) that could estimate the attention with the help of two electrodes: one for the true signal estimation and other for the noise estimation. The algorithm is tested using EEG signals collected from ten subjects and its performance is compared against the state-of-the-art LS algorithm.