Domain Adaptation for Attention Steering

Abstract: A major problem in the development of intelligent hearing aids is often referred to as the cocktail party problem. It describes the remarkable ability of the human brain of filter out unwanted sounds in a noisy environment, while focusing on a single talker or conversation. Without the ability to select and enhance a specific sound source of choice while suppressing the background, the hearing aids generally amplify the volume of everyone in the environment. The problem of knowing which speaker to enhance is unsolved and most people with hearing aids still experience discomfort in noisy environments. This thesis uses EEG data from real-life scenarios where the subjects for each trial listened to one female voice and one male voice at the same time while giving attention to one of the speech streams. The stories were simulated to come from a distance of 2.4m in a direction of ±60° from the listener. Due to both instrumental and human factors, data from different subjects will differ and it is not possible to create a classifier which works on all data. It is said that the data from each subject lives in different domains, and they need to be transported to the same domain in order to be classified together. The transportation is called domain adaptation, and this thesis have used and compared two domain adaptation methods: Parallel transport and Optimal transport. Two different classification problems are considered in this thesis: attention to male voice vs female voice and attention to left side vs right side. The classification accuracy differed greatly depending on which data was used. Generally, the results were better for male/female separation which almost always gave successful results, and the highest classification accuracy reached 95%. Transportation of several subjects for the left/right separation problem did not give results above the level of chance, however the best classification accuracy reached above 93% which is considered a successful result.