Spectrum Weighted HRTF Based Sound Localization

In the framework of humanoid robotics it’s of great importance studying and developing computational techniques that enrich robot perception and its interaction with the surrounding environment. The most important cues for the estimation of sound source azimuth are interaural phase differences (IPD), interaural time differences (ITD) and interaural level differences (ILD) between the binaural signals. In this paper we present a method for the recognition of the direction of a sound located on the azimuthal plane (i.e. the plane containing the interaural axis). The proposed method is based on a spectrum weighted comparison between ILD’s and IPD’s extracted from microphones located at the ears and a set of stored cues; these cues where previously measured and stored in a database in the form of a Data Lookup Table. While the direct lookup in the table of the stored cues suffers from the presence of both ambient noise and reverberation, as usual in real environments, the proposed method, exploiting the overall shape of the actual frequency spectrum of the signal, both its phase and modulus, reduces dramatically errors in the localization. In the paper we give also the experimental evidence that such method improves greatly the usual HRTF based identification methods.