Exploiting joint sparsity for far-field microphone array sound source localization

Abstract

The presence of far-field noise and reverberation poses significant challenges to the conventional microphone array sound source localization approaches. Consider the sparsity contained in the source direction vector, source localization can be transformed into a compressed sensing (CS) problem by constructing the redundancy frequency domain room impulse response (RIR) matrix as CS measurement matrix. In this paper a new sparse recovery model is derived by decomposing the RIR into delay response term and reverberation response term to facilitate reverberation mitigation via frequency domain accumulation. Furthermore, as the source direction vector of adjacent speech frames tends to exhibit similar sparse pattern, namely, the direction of source can be assumed to keep static within this short period, thus there exists substantial correlation of spatial sparsity among adjacent speech frames. In this paper, under the framework of distributed compressed sensing (DCS), multiple source direction vectors are treated as sparse solutions with common spatial support to derive a joint sparse recovery algorithm for far-field source localization. The experimental results obtained in the context of a uniform circle array (UCA) show that the proposed algorithm is capable of yielding better estimation performance compared with the traditional algorithms.