Real-Time Independent Vector Analysis for Convolutive Blind Source Separation

Utilizing dependence over frequencies has shown significant excellence in tackling the frequency-domain blind source separation (BSS), which is also referred to as independent vector analysis (IVA). The IVA method then runs in offline batch processing, which is not well applicable to real-time systems. This paper proposes real-time BSS methods corresponding to that model. First, we derive online algorithms under some assumptions. Then, in order to improve the performance and convergence properties, a modified gradient with nonholonomic constraint and a gradient normalization method are proposed. The convergence speed is improved by the gradient normalization. The gradient with nonholonomic constraint shows better performances, although it has less computational complexity. In addition, the proposed method has a simpler structure than any other existing methods and runs in fully online mode. Thus, it requires sufficiently less computations and memories. Based on these benefits, the algorithm is implemented in a real-time embedded system. The experimental results confirm effectiveness of the proposed method with both simulated data and real recordings.