| Abstract | Modern audio systems are typically equipped with several user-adjustable parameters unfamiliar to most listeners. In order to obtain an optimal system setting, the listener is nevertheless forced to perform high-dimensional optimization with respect to the user's own objective. In the present paper, a general inter-active framework for performing robust personalization of such audio systems is proposed, which addresses the problems associated with traditional trial and error methods. The framework builds on Bayesian Gaussian process regression in which the extit{belief} about the user's extit{objective function} is updated sequentially. The setting to be evaluated in a given trial is then carefully selected by sequential experimental design. A modified Gaussian process model is suggested that assumes adjacent parameters to be correlated, which shows better modeling abilities compared to a standard model. We further demonstrate the framework in an interactive loop, where twelve test subjects obtain a personalized setting in a five-band constant-Q equalizer. The proposed approach is able to find a significantly better solution than obtained with random experimentation. |