Abstract

Multiple sound source localization is an important application in speech processing. In this paper, a cuboids nested microphone array (CuNMA) is proposed for sound acquisition. Also, the spatial aliasing is eliminated by the use of this array. Then, the subband processing is proposed based on the GammaTone filter bank. In the next, the generalized eigenvalue decomposition (GEVD) algorithm is implemented on all microphone pairs of CuNMA and for each obtained subband of the GammaTone filter bank. In each subband, the standard deviation (SD) is calculated for all direction of arrival (DOA) estimations, and the subbands with improper information are eliminated. Then, the K-means clustering with silhouette criteria are implemented on all DOAs for estimating the number of speakers and to allocate the related DOAs for each cluster. The proposed method is compared with steered response power-phase transform (SRP-PHAT), Geometric Projection, and spectral source model-deep neural network (SSM-DNN) on simulated data in noisy and reverberant conditions, which the results show the superiority of the proposed method in comparison with other previous works.