Dynamic blind source separation based on source-direction prediction

Deterministic techniques are based on the source-directions and multipath characteristics of the reverberant environment for different source signals. However, searching for the desired directions of the time-block sequence of an acoustic signal is time consuming, and existing deterministic methods rarely consider the motion properties of the acoustic source. In this paper, a dynamic source-direction prediction method for real-time blind convolutive mixtures based on a Kalman filter is proposed. First, the convolutive mixture signals captured by the coincident array geometry are formulated, and the relationship between source-direction and source separation is developed. Second, motion prediction based on a Kalman filter is theoretically analyzed, and the motion of a source is modeled as a noise-driven position integrator with enough samples. Then, a dynamic source-direction prediction method for real-time blind source separation based on a Kalman filter is proposed to predict the directions of a time sequential signal. Combined with the local direction searching method, our proposed method has a self-correction ability according to the three-sigma rule. Finally, extensive experiments are performed with three-source convolutive mixtures of speeches in English and Chinese, whose direction varies in linear and nonlinear motions. The signal-to-distortion and signal-to-interference of the separated signals are calculated, and the experimental results demonstrate the feasibility and validity of the proposed method.