If I had a vector containing the spectral centroid, spread, skewness and kurtosis (all normalised) of an audio clip and found the distance between it to the same kind of vector of another audio clip. Would this give me a very crude estimation of how likely each audio clip would mask each other? Bear in mind I'm aware of how loud each audio clip has to be. I'm just trying to see if their spectrums would interfere with each other. Or could I possibly just find the average over time the amount of energy in each FFT bin to build a spectral profile of audio clip, so it would be essentially be like finding the distance between two 1024 value vectors (normalised obviously)
By looking at the spectrum's probability law you are completly missing the temporal information in the two signal you wish to compare thus it would be a bad predictor. Imagine you analyse two 2048 pts signal, one is null during the last 1024 points whereas the other one is null during the first 1024 pts, they could have the exact same spectrum but they won't interfere each other at all. This is the whole purpose of time-frequency analysis, unless you have some a priori knowledge, you can't avoid using a short time fft.