We have profitably used many of the known properties of the inner-ear in our spectral models. For example, the peak-dominance of audio perception matches well with the ``unreasonably effective'' sinusoidal model. Similarly, as MPEG audio and S+N+T models show, we can inaudibly eliminate over of the information in a typical sound, on average.
An interesting observation from the field of neuroscience is the following [#!GardnerAndMagnesco06!#]:
``... most neurons in the primary auditory cortex A1 are silent most of the time ...''This experimental fact indicates the existence of a much sparser high-level model for sound in the brain. We know that the cochlea of the inner ear is a kind of real-time spectrum analyzer. The question becomes how is the ``ear's spectrogram'' processed and represented at higher levels of audition, and how do we devise efficient algorithms for achieving comparable results?