Real world sound localization by using the binaural model

Abstract

In this study, we proposed a sound localization method by using the binaural model for a real environment, in which the reflected sound and background noise is present. In order to estimate of horizontal angle, the reduction method of the initial reflected sound that combining the modified Thomposon-τ method and the modified sub-band peak hold processing, was proposed and its effectiveness has been performed. In addition, to create the ESR database of vertical angle, the modified Thompson-τ method is used for elimination of the abnormal value. In the estimation of the vertical angle, a new method that is called wins calculation method is proposed and it is strong in abrupt noise, without using the traditional cross-correlation method. The experiments were carried by using /a/ sound in real environment with 45[dB] background noise and 300[ms] reverberation time. As the results obtained that the estimate correct rate of the horizontal angle is 94.6%, the correct rate of the vertical angle is next 98.6%, and the validity of the proposed method was confirmed.