This paper describes the design and the experimental validation of a preprocessing block for an automatic modulation classifier. The preprocessing block performs two main tasks. The former consists of dividing the spectrum in sub-bands, while the latter has the aim of sensing the signal presence, in each subband. The signal sensing is provided by a method, based on the eigenvalues of the covariance matrix of the acquired record. Several experimental tests with simulated and actual signals, for different values of the Signal to oise Ratio, have been performed showing very high detection percentages.