Use of prior information in acoustic inference

Abstract

Often in scientific and engineering applications, we possess prior information about the specific problem under investigation. In acoustics, this prior information may consist of knowledge about the design and arrangement of our detectors, the behavior of the sound sources, and the laws of physics describing sound propagation through an intervening medium. In this talk I explore the use of prior information for use in Bayesian inference in acoustic applications. This prior information can be either encoded in the signal model, which is represented in the Bayesian formulation by the likelihood function, or in the prior probabilities, which weight the particular values of the signal model parameters. I give examples of the use of such prior information by focusing on the specific problems of source separation and localization. By choosing to focus on the appropriate model parameters one can change a source separation problem into a source localization problem and vice versa, thus changing the very nature of the acoustic inference.