Comment by kyzyl

The tools to do this exist. It's usually called 'blind source separation', as in "What are the N distinct audio signals which sum up to best explain a given compound signal, without knowing the possible source signals ahead of time." Usually it's done with some sort of matrix factorization, Principal Component Analysis, and/or Independent Component Analysis. It's also used for non-audio signals, like pulling the discrete firings out of noisy EEG signals. It's definitely not a foolproof solution but in a lot of applications it can get you going, at least.