Abstract: A method for separating a signal mix composed of the additive superposition of M known functions that are weighted with M unknown weighting factors, whereby all or some of the functions are chronologically and/or spectrally not disjunctive. The signal mix including a potential noise signal is supplied to a signal evaluation device that, by filtering, forms true-to-expectation estimated values for the unknown weighting factors. For the first time, this provides a true-to-expectation separation of a signal mix composed of non-orthogonal functions. An apparatus for the implementation of the method forms linear combinations of the correlation products from the input signal and from the functions to be separated. Instead of the functions to be separated, versions of these functions modified according to the properties of the superimposed noise signal can also be used in the formation of the correlation products.