LSTMs for EEG-based Auditory Attention Decoding

The cocktail party problem occurs when hearing aid users are in an environment with more than one speaker. For these people, it is difficult to follow a particular speaker. One possible solution would be to use electroencephalography signals recorded by the hearing aid to identify which speaker to listen to and amplify it. To improve this detection, we developed a long-short-term-memory-based neural network architecture that detects which of two speakers the person is currently paying attention to. The model was tested with two different preprocessing pipelines on two publicly available datasets. The network achieved 92% accuracy with an input window length of 0.25 seconds, the highest value for the dataset used.