A half-frequency domain approach for convolutive blind source separation based on Kullback-Leibler divergence

Abstract

This paper studies the problem of blind separation of convolutive mixtures by means of Kullback-Leibler divergence in frequency domain. Unlike exiting approaches, an integrated objective function is defined in frequency-domain with time-domain parameters as the variables. The permutation problem is avoided through the frequency-domain integration and time-domain optimization. Simulation results show that the algorithm is valid and of high performance for the separation of real-world recorded convolutive mixtures.