Decoding Auditory Attention Based On EEG Signals

Humans show the remarkable ability to selectively attend to one speaker in a noisy social environment.However,the existing hearing aids only have simple functions,which cannot intelligently assist the subjects with hearing loss to follow the sound of interest in cocktail party scenarios.As a result,hearing-impaired people often struggle in daily communication and suffer from low quality of life.Benefiting from the research progress in related areas such as psychoacoustic,biophysiological,and neuroscience regarding the brain activity of auditory attention,researchers are inspired to develop a computational model to detect the attention activities manifested in brain signals.Auditory attention detection(AAD)based on brain signals sheds light on the implementation of neuro-steered hearing aids,which is also important progress in artificial intelligence research.However,current AAD approaches show relatively low decoding performance and are particularly problematic in low-latency settings.One possible reason is that electroencephalography(EEG)signals feature high complexity and low signal-to-noise ratio.Previous AAD models do not fully exploit the multivariate information of EEG,which leads to limited decoding performance.Meanwhile,previous studies ignore the physiological mechanisms of the auditory cognitive process in the human brain and oversimply the AAD task.Moreover,the interpretation of traditional AAD models has been difficult to achieve because of the inherent series of nonlinear transformations,which is also known as the black box problem.To address these limitations,this study firstly proposes a bio-inspired AAD model based on EEG signal and dynamic feature extraction,which aims at fast and accurate decoding of auditory attention.Specifically,this study proposes a novel feature representation method to dynamically learn EEG features through frequency-wise,a channel-wise,and time-wise attention mechanisms.With these neural attention mechanisms,our model is capable of focusing on the informative features related to the AAD tasks and extracting more discriminative EEG features.Moreover,this study develops a decoder for detecting the target speaker as well as a decoder for detecting the locus of the target speaker,respectively.The feasibility and effectiveness of the proposed decoders are validated by comprehensive experiments on three databases and comparative studies.In addition,the interpretability of the proposed AAD models is realized by using data visualization,combined with the basic theory of auditory cognition research in the brain.To sum up,the high accuracy on very short data segments achieved by the proposed model is a major step forward towards practical neuro-steered hearing devices.To the best of our knowledge,this is the first AAD model that takes the cognitive mechanism of the human brain into consideration,which improves the generalization ability and interpretability of the model.The obtained results are consistent with existing neuroscientific findings,which show potential benefits for further exploring the auditory cognitive mechanism of the brain.