A trade-off pre-whitening method based on piecewise bilinear prediction for speaker localization

Speaker localization plays a fundamental role in numerous acoustic application systems. The approach based on maximizing the steered response power (SRP) of a beamformer is an important and practical method. An SRP method with bilinear prediction-based trade-off pre-whitening was developed to enhance the robustness over noise and reverberation. However, this method is of high computational complexity. This paper proposes an improved trade-off pre-whitening method to pre-filter microphone signals. The coefficient vector of the linear predictor that is used to model the speech signals is first divided into two segments of equal length. Each segment is then decomposed through bilinear prediction to establish a set of convex-constrained linear prediction sub-models for pre-whitening microphone signals. Thanks to the piecewise bilinear form of the predictor, the high-dimensional correlation matrix is converted into a set of low-dimensional correlation matrices, significantly reducing the computational complexity of the pre-whitening algorithm. Simulation results demonstrate the effectiveness and computational efficiency of the proposed method.