Auditory Scene Analysis: Computational Models

Listeners have to make sense of a complex acoustic world containing overlapping sound sources that must be organized into individual auditory objects. Computational auditory scene analysis concerns the use of algorithms inspired by human sound perception whose aim is to extract properties of constituent sound sources in a complexmixture. Starting with representations based on models of how sound is processed in the peripheral auditory system, typical computational auditory scene analysis techniques function by decomposing the mixture into components followed by selective recomposition into groups of components that appear to emanate from a single source. Grouping processes can be informed by information in the signal itself or by the use of prior statistical models of sound sources. This article outlines some of the principal signal decompositions used in models of auditory grouping and goes on to describe a decoder that combines both signaland modeldriven grouping processes.