Convolutive Blind Source Separation with Independent Vector Analysis and Beamforming

This paper considers the cocktail party problem for determined setup with convolutive mixtures. The proposed framework combines independent vector analysis with beamforming in order to improve the performance. Our approach has the advantage of not requiring a permutation alignment algorithm or accurate knowledge of all source locations. The performance of the proposed framework is evaluated and compared to conventional techniques in terms of accuracy of recovery, SIR gain and performance under different reverberation times. Simulation results demonstrate the clear superiority of the method especially in reducing the destructive correlation as well as offering a more stable performance with varying reverberation times. Furthermore, results show beamforming can also help reduce the convergence time of the separation algorithm.