A Robust Dual-Microphone Speech Source Localization Algorithm for Reverberant Environments

Speech source localization (SSL) using a microphone array \naims to estimate the direction-of-arrival (DOA) of the speech \nsource. However, its performance often degrades rapidly in reverberant \nenvironments. In this paper, a novel dual-microphone \nSSL algorithm is proposed to address this problem. First, the \ntime-frequency regions dominated by direct sound are extracted \nby tracking the envelopes of speech, reverberation and background \nnoise. The time-difference-of-arrival (TDOA) is then \nestimated by considering only these reliable regions. Second, \na bin-wise de-aliasing strategy is introduced to make better use \nof the DOA information carried at high frequencies, where the \nspatial resolution is higher and there is typically less corruption \nby diffuse noise. Our experiments show that when compared \nwith other widely-used algorithms, the proposed algorithm produces \nmore reliable performance in realistic reverberant environments.