Solving the Cocktail Party Problem: Spectral Estimation and Linear Modelling

By measuring brain activity, through techniques such as electroencephalography (EEG), it is possible to decode which sound source a person is listening to, called auditory attention decoding (AAD). This can either be done investigating the relation between speech sources and corresponding brain responses over time, or by discrimi-natively estimating directions to which auditory attention is focused. Spectral, temporal and spatial information are all useful and each essential for understanding how the brain processes sounds in a multi-talker scenario. Key challenges with EEG analysis are high levels of noise from various sources, as well as utilizing methods that infer onto the processing happening in the brain. Therefore, the work part of this thesis focuses on linear and fairly non-complex methods. This thesis explores spectral estimation based methods and linear modelling methods and their application to AAD. The linear correlation measure of coherence is investigated and improved for use in EEG and AAD, showing that it can differ between attended speech and ignored speech. The commonly applied method of common spatial patterns (CSP) within EEG-data is employed specifically for AAD. We are able to show how different CSP algorithms perform within the field of AAD, and that performance for CSP carries over from decoding auditory attention of individuals with normal hearing compared to individuals with hearing impairment. Independent Component Analysis-based (ICA) methods of removing noise components of EEG data are evaluated for AAD on a dataset with participants hearing impaired. Automatic noise cleaning methods are shown to perform equally as well as the traditional manual method on the given dataset. Finally, a phase estimation technique for transient components based on spectrogram reassignment is developed, which can estimate phase difference of signal components in multi-channel measurements such as EEG. Using the methods described, it is possible to draw interesting conclusions in the field of AAD. However, future work entails further improvement and exploration of useful methods for analysis of the system that is the hearing brain.