Improving Classification of Auditory Attention using Optimized Multitapers and Machine Learning

Sammanfattning: The cocktail-party problem relates to the challenge of separating a single sound source from a noisy and crowded background based on the listener’s attention. Earlier work has been done in order to use the relations between the sounds heard and the following brain response to develop a model that can accurately classify which sound an individual listener is listening to. The model deals with an reconstruction based approach, where the brain response is used to reconstruct the speech stimuli. Pearson’s correlation coefficient is then used to decide which speech is attended to. This model has been evaluated on a dataset with low to no noise, but in this thesis however, background noise at different levels have been added further complicating the problem. The objective of this thesis is to further improve on this classification model, and to evaluate whether the background noise has an effect or not. The first proposed area of improvement is to use different methods of estimating the cepstra, for improved feature extraction. In particular, it shall be investigated if using optimally weighted multitapers notably improves results. The second proposed area for improvement is using more sophisticated reconstruction algorithms, using instead Machine Learning techniques, in particular support vector regression and neural networks.