Multi-sources separation for sound source localization

Spatial localization given measurements of the sound field at a microphone array is a computationally challenging task. While big progress has been achieved in Azimuth localization, little attention has been paid to localization in distance. This paper presents a multi-sources localization model based on Time Delay of Arrival (TDOA) using only two microphones. The proposed method requires estimation of mixing parameters, which are informative about speakers’ positions. The mixing parameters are estimated by determining the slope of a scatter plot of the separated sound source from the time delayed mixed speech captured by each microphone. In the determined context, experiments on BSS-eval dataset show that the proposed model achieves distance accuracy of more than 92% with a less than 1,8m distance localization error.