A method includes receiving at least a first input audio channel and a second input audio channel, and using an inter-channel prediction model to form at least one inter-channel parameter. The first and second input audio channels represent a spatial audio image of an acoustic space. The inter-channel prediction model is a linear prediction model representing a predicted sample of the first input audio channel using a weighted linear combination of samples of the second input audio channel. An apparatus for practicing the method and a corresponding computer program product are also disclosed.