Sound source localization based on Gaussian mixture models using ITD with diffuseness mask in a neck-band microphone array module

Sound source localization using multiple array microphones is recently required in various fields, especially in wearable devices such as neck-bands that recognize the direction of dangerous sound sources for the hearing impaired. In this paper, we propose an algorithm based on Gaussian mixture models (GMMs) using the interaural time difference (ITD) as a method for sound source localization in neck-band devices. In addition, we apply a diffuseness mask for robust sound source localization in diffuse noise environments. The ITD feature to be learned in GMMs is extracted through generalized cross-correlation with phase transform (GCC-PHAT), and the diffuseness mask is estimated using coherent-to-diffuse power ratio (CDR) based on the spatial coherence between microphone observations to select time-frequency (t-f) bins providing ITDs robust to diffuse noise. Experimental results using real recorded data show that the GMM-based sound source localization with the diffuseness mask is robust in diffuse noise environments.