Localization of Closely-Spaced Speech Sources Based on Small Microphone Arrays

Abstract

Indoor source localization based on a microphone array is still a difficult problem especially for closely-spaced sources in strong reverberation environments, in such cases a large number of microphones are often utilized. Considering the industrial design and cost, source localization using a small microphone array in complex conditions is worthy of investigation. In this paper, we propose a small-size array based localization method under the condition of close proximity and strong reverberation. The essence of the proposed method lies in the detection of single source time-frequency (TF) points (SSPs) of each source, which simplifies the problem of multisource localization in close proximity into several single-source localization ones. The detected SSPs are then used for precise source localization in a sparse Bayesian framework, wherein the reflection coefficient can be simultaneously estimated by implementing a Bayesian dictionary learning. Experimental results confirm the effectiveness of our method in locating spatially-close sources compared with some state-of-the-art methods.