Monaural Separation of Dependent Audio Sources Based on a Generalized Wiener Filter

Abstract

This paper presents a two-stage approach for single- channel separation of dependent audio sources. The proposed algorithm is developed in the Bayesian framework and designed for general audio signals. In the first stage of the algorithm, the joint distribution of discrete Fourier transform (DFT) coefficients of the dependent sources is modeled by complex Gaussian mixture models in the frequency domain from samples of individual sources to capture the properties of the sources and their correlation. During the second stage, the mixture is separated through a generalized Wiener filter, which takes correlation term and local stationarity into account. The performance of the algorithm is tested on real audio signals. The results show that the proposed algorithm works very well when the dependent sources have comparable variances and linear correlation.