A hybrid equalizer using a linear combiner and a recurrent neural network is presented. It characterizes itself for being adaptive and presenting: 1) a worst case performance very close to the best of the substructures that composes it, being better that each one of them in critical situations; 2) a computational complexity that makes its implementation feasible; and, 3) a good performance in difficult environments as, for example, channels with non-minimum phase, spectral nulls or non-linearities. Adaptation of coefficients is done using both the LMS and RTRL algorithms. Simulations illustrate the good performance of the proposed equalizer.