Single channel source separation using static and dynamic features in the power domain

Abstract

In this paper we address the problem of separating the sound sources composing complex sound mixtures using a single microphone. The a-priori information of static and delta power of each source is represented by Gaussian mixture models (GMMs) and incorporated into a full posterior probability density function. We present a unified probabilistic framework that integrates the a-priori information of the power and the delta power of the sources and we derive a closed-form approximate minimum mean square error (MMSE) estimator of the audio sources. The experimental part evaluates our approach on mixtures of real environmental sounds in scenarios that involve speakers talking in a music background. Comprehensive experiments clarify the importance of incorporating delta in the separation process by presenting separation results using the static only and the joint static and delta a-priori models.