From sound waves to locations : computational models for sound source localization in the early auditory pathway

Abstract

Audition is our only far-reach omni-directional sense and it plays an important role for survival. Detecting a prey hidden in bushes or a predator sneaking up behind us can be crucial. Our auditory system facilitates us to not just detect such sound sources but also precisely localize them. This localization is not trivial, but requires intensive processing steps along the auditory pathway. These steps include the binaural integration of signals from the left and right ear and the identification of systematic modifications in the perceived spectrum of a sound. In particular, to localize sound sources in the horizontal plane the interaural time (ITD) and level difference (ILD) are computed. However, these cues create ambiguous information in the vertical plane. Thus, elevation-dependent spectral cues created by the shape of the ear are identified to localize sounds in the vertical plane. Together with the ITD and ILD, these cues provide a 2D representation of auditory space. This space is comprised of internally computed auditory cues and therefore needs to be calibrated to create an association with the external world. Such a calibration is realized with a visual guidance signal to further create a common frame of reference and a means for multisensory integration. Computational models that describe all these processing steps are currently either missing or merely specialized on a single stage of the processing pathway.

In this thesis, the auditory pathway for localizing sound \\ sources is investigated by developing canonical computational neural network models of consecutive processing stages. In particular, a model for ILD computation in the lateral superior olive is presented which incorporates a synaptic adaptation mechanism that leads to improved spatial localization of sounds in the horizontal plane. Furthermore, a novel model of spectral cue processing for vertical sound source localization is introduced. This model suggests that localization in the vertical plane is fundamentally binaural but allows for monaural sound source localization. A model of the inferior colliculus describes how cues for horizontal and vertical localization are aligned with a visual teacher signal to provide a common frame of reference. Finally, a multisensory integration model of audio-visual inputs of the superior colliculus demonstrates the importance of cortical feedback for top-down modulation of sensory input. The analysis provides insights on its effect on the Bayesian optimal fusion of multisensory input streams.

Model conception and simulation experiments in this thesis investigate underlying neural principles and provide new hypotheses for auditory sound source localization.