Abstract

© 2019 Division of Signal Processing and Electronic Systems, Poznan University of Technology (DSPES PUT).Multiple sound source localization is one of the most important applications in speech processing. The challenge in localization and tracking algorithms is to have better accuracy in noisy and reverberant environments. In the proposed method in this paper, a Quasi-Spherical Nested Microphone Array (QS-NMA) is suggested to eliminate the spatial aliasing and to be applicable for 3D sound source localization. In addition, the microphone signals related to QS-NMA are divided to different subbands by GammaTone filter bank based on the speech spectrum components. The subband processing is considered due to the W-Disjoint Orthogonality (W-DO) of speech signal specially in low frequencies. Then, the modified steered response power (SRP) is implemented based on the specific microphones of QS-NMA and subband signals. The modified SRP method is combined by ML and PHAT weighted functions adaptively and the peak positions of the modified SRP are extracted based on the number of speakers. This process is implemented on all subbands and the final histogram is calculated by combination of histograms for each subband. The 3D positions of all speakers are estimated by peak selections of the final histogram based on the number of speakers. The Proposed system is evaluated on different noisy and reverberant conditions and the superiority of the method is presented in comparison with other previous works. This system by using of QS-NMA localizes speakers in different directions with the same probability for speaker's positions in indoor conditions.