Concept
A song's spectrograph displays all of the different sounds in the song graphically (by plotting the amplitudes of all audible frequencies, probably on a logarithmic scale of some kind).
Unfortauntely, it's difficult to tell the different sounds apart. Frequencies points alone a very poor visualization. One reason is that all the different sounds in the song are combined in one complex pattern. You don't know if the 440hz waveform with amplitude x came from a piano, or a flute, or a guitar, or all three at the same time.
There just isn't enough space or detail for you to visually distinguish one sound from another.
However, if the pattern of each sound's temporal morphology could be recognized by an algorithm, the individual sounds could be separated out and displayed in different layers. Each plane could be coloured differently, and possibly sub-coloured to express the morphological attributes that the system recognizes.
This would look really cool.
Of secondary benefit is the ability to manipulate the isolated layers of the song. This would have analytical uses -- like filtering sounds -- or creative uses -- like selectively applying interesting effects or edits to the layers.
Example Spectrographs
dj grazhoppa - milky rmx
simian mobile disco
announcer
cheesy organ music
daft punk (technologic)
Resources
DNA does something similar to what's needed here, but the goal is different; we don't need to be able to select and transform the individual entities in the sound layer, we just want to quickly mask a single layer so that its sound data can be easily piped around.