A processing unit for processing a multi-channel audio signal has a delay element (40) for delaying an FM sum signal (sum) and a converter arrangement (T) for converting an FM difference signal (diff) and a noise signal (diffnoise) to the frequency domain. Frequency-based noise suppression is used to derive a de-noised frequency-domain difference signal using a gain function which is limited to a maximal and a minimal value. This is then converted to the time domain, and the first and second audio signals are calculated from a delayed sum signal (sum2) and the de-noised time domain difference signal (diff2).