Sound Source Localization by Proposed Subband Adaptive GEVD Algorithm Based on GammaTone Filter Bank in Undesirable Acoustical Conditions

摘要

Sound source localization methods are implemented with different algorithms. Some methods are based on received energy to microphones and some other methods estimate the Time Difference Of Arrival (TDOA) of sound sources. The energy-based methods have high computational complexity but the Time Delay Estimation (TDE)-based methods have low accuracy in undesirable acoustical conditions. Adaptive Generalized EigenValue Decomposition (GEVD) algorithm is a method for TDE that has appropriate accuracy in noisy conditions but it does not have an acceptable performance in noisy-reverberant scenarios. The proposed method in this paper is subband adaptive GEVD based on GammaTone filter bank for TDE. Since the speech signal information is different in frequency bands, then it is necessary to present a method to concentrate on low frequency components of speech signal to have better accuracy for sound source localization. In proposed method, firstly the two microphone signals are divided to different subbands with GammaTone filter bank. This filter bank is designed based on the human auditory. Then, the GEVD function is implemented on these subbands information. Finally, the output of GEVD function are weighted and combined based on the frequency spectrum energy in different subbands. The experiments on noisy and reverberant scenarios show the superiority of proposed method in comparison with GEVD algorithm in different scenarios. Although proposed method has more computational complexity because of subband processing, but the improvement in accuracy compensate this complexity.