Figure 4.

Generative model explaining how the input spectrum results in an elevation percept. The sensory spectrum is a convolution of the sound with the HRTF of its associated elevation, ε^∗. Different frequency bands around the notch area are weighted differently, depending on the strength and reliability of their spatial information content, resulting in a weighted sensory spectrum, S_W(f, ε^∗). The latter is cross-correlated with stored information about all HRTFs, resulting, after rectification, in a likelihood distribution of potential source locations, L(ε|ε^∗). Subsequently, Bayesian inference weighs the sensory evidence against its internal prior (taken around straight ahead), resulting in a more precise posterior. Finally, the decision stage selects the maximum of the posterior (MAP), to provide an optimal estimate for the perceived elevation angle, ε_P. Blue labels indicate putative neural stages, described in the text.