A probabilistic model for robust acoustic localization based on an auditory front-end

Although extensive research has been done in the field of localization, the degrading effect of reverberation and the presence of multiple sources on localization performance has remained a major issue. The classical approach to localize an acoustic source in the horizontal space is to search for the main peak in the cross-correlation function, which corresponds to the interaural time difference (ITD) between both ears. Apart from ITD, the interaural level difference (ILD) can contribute to localization, especially at higher frequencies where the wavelength becomes smaller than the diameter of the head, leading to ambiguous ITD information. Motivated by the robust localization performance of the human auditory system, its peripheral stage is used as a front-end for binaural cue extraction. The interdependency of ITD and ILD on azimuth is a complex pattern that depends also on the room acoustics and is therefore learned by azimuth-dependent Gaussian mixture models. Multiconditional training is performed to incorporate the spread of the binaural features caused by multiple sources and the effect of reverberation. The trained localization model outperforms state-of-the-art localization techniques in simulated adverse acoustic conditions. Furthermore, the model is capable of generalizing to changes in the simulated room absorption and to unknown source/receiver combinations