Robust Localization of Multiple Speech Sources Based on Time Difference of Arrival in Real Environments for Binaural Robot Audition

This paper presents a multisource speech localization method based on the generalized cross-correlation (GCC) method weighted by the phase transform (PHAT) for binaural robot audition. The direction-of-arrival (DOA) estimation based on the GCC-PHAT method was extended to enable simultaneous multiple DOA estimations with a signal-to-noise ratio (SNR)-based weighting function. The standard K-means clustering algorithm was improved for the purpose of multisource speech localization by adding two additional steps that increase the number of clusters automatically and eliminate clusters containing incorrect DOA estimations. Experiments conducted on the SIG-2 humanoid robot in real environments show that our method can localize multiple speech sources in real-time with localization error below 5.96°.