Part of Advances in Neural Information Processing Systems 9 (NIPS 1996)
Hagai Attias, Christoph Schreiner
In order to process incoming sounds efficiently, it is advantageous for the auditory system to be adapted to the statistical structure of natural auditory scenes. As a first step in investigating the relation between the system and its inputs, we study low-order statistical properties in several sound ensembles using a filter bank analysis. Focusing on the amplitude and phase in different frequency bands, we find simple parametric descriptions for their distribution and power spectrum that are valid for very different types of sounds. In particular, the amplitude distribution has an exponential tail and its power spectrum exhibits a modified power-law behavior, which is manifested by self-similarity and long-range temporal cor(cid:173) relations. Furthermore, the statistics for different bands within a given ensemble are virtually identical, suggesting translation in(cid:173) variance along the cochlear axis. These results show that natural sounds are highly redundant, and have possible implications to the neural code used by the auditory system.