Sound texture recognition through dynamical systems modeling of empirical mode decomposition.

This paper describes a system for modeling, recognizing, and classifying sound textures. The described system translates contemporary approaches from video texture analysis, creating a unique approach in the realm of audio and music. The signal is first represented as a set of mode functions by way of the Empirical Mode Decomposition technique for time/frequency analysis, before expressing the dynamics of these modes as a linear dynamical system (LDS). Both linear and nonlinear techniques are utilized in order to learn the system dynamics, which leads to a successful distinction between unique classes of textures. Five classes of sounds comprised a data set, consisting of crackling fire, typewriter action, rainstorms, carbonated beverages, and crowd applause, drawing on a variety of source recordings. Based on this data set the system achieved a classification accuracy of 90%, which outperformed both a Mel-Frequency Cepstral Coefficient based LDS-modeling approach from the literature, as well as one based on a standard Gaussian Mixture Model classifier.