One-channel separation and recognition of mixtures of environmental sounds: The case of bird-song classification in composite soundscenes

Abstract

In this paper, we address the problem of automatic separations and recognition of bird vocalizations in a mixture of other environmental sounds using a single microphone. We present a novel single-channel audio separation method that trains statistical models of the sources to perform the separation on a feature space of low dimensionality and we lead the separated streams to automatic recognition engines that recognize general sound events. The experimental part tests and evaluates the system on mixtures of bird and insects songs as well as dog barks and other environmental sounds. © 2008 Springer-Verlag Berlin Heidelberg.