A BSS-based approach for localization of simultaneous speakers in reverberant conditions

Abstract

In this paper, we address the localization of simultaneous speakers by means of Blind Source Separation (BSS) based algorithms. Considering BSS demixing filters as some blind null beamformer and producing an acoustical map from them, source localization can be achieved by identifying the local minima of this acoustical map. To improve the performance of this method in reverberant environments, we have proposed to replace the demixing filter with one, corresponding to the direct path only. This is done by keeping only the largest coefficient in each demixing filter and neglecting the other coefficients. Besides, the proposed method reduces the computational complexity. To further improve the computational efficiency of the localization method, we have also proposed the limitation of the frequency range within averaging procedure. The experimental results demonstrate improved accuracy and efficiency of the proposed method in the localization of multiple simultaneous sound sources in reverberant environments.