Modeling human sound localization with hierarchical neural networks

Artificial neural networks were trained to identify the location of sound sources using head related transfer functions (HRTFs). The simulated signals were filtered clicks presented from virtual speakers placed at 15 degree steps in azimuth and 18 degree steps in elevation. After signals were passed through the HRTFs, quarter-octave spectra were computed. The inputs to the networks were either monaural spectra or binaural difference spectra. In some cases a broadband cross-correlation term was also provided. Backpropagation was used to train the networks. Separate networks were trained for each combination of spectral information. In some cases the networks achieved performance comparable to that of human observers in both accuracy and number of front-back reversals. With a hierarchy of neural networks accuracy better than human performance was obtained