We propose a novel algorithm for the identification of a Multi-Input-Multi-Output (MIMO) system. Instead of using “classical” high-order statistics, the mixing system is estimated directly from the empirical Hessian matrices of the second generalized characteristic function (GCF) at several preselected “processing points”. An approximate joint-diagonalization scheme is applied to the transformed set of matrices in the frequency domain. This yields a set of estimated frequency response matrices, which are transformed back into the time domain after resolving frequency-dependent phase and permutation ambiguities. The algorithm’s performance depends on the choice of processing points, yet compares favorably with other algorithms, especially at moderate SNR conditions.