Single channel enumeration and recognition of an unknown and time-varying number of sound sources

In this paper we address the problem of on-line enumeration and classification of an audio mix of spectrally overlapping sound sources using a single microphone. We assume that the mix is composed of a subset of realizations of sound sources that belong to an a-priori known set of classes. Each class is represented by a Gaussian mixture model (GMM) probabilistic density function trained from available recordings of each sound class. The approach is based on forming multiple hypotheses on the composition of the mix set that are propagated through time and tested against the likelihood of having produced the audio mix. The likelihood acts as a switch that deletes or reinforces hypotheses. The hypothesis generation and evaluation process is set under a unifying particle filtering framework that estimates the cardinality of the set of sources composing the mix under the maximum a-posteriori (MAP) criterion as well as the sources themselves. The experimental part tests and evaluates the algorithm on real composite environmental soundscenes and on simulations involving rapid changes of cardinality of a set of Gaussian sources.