A new method of reducing the coefficient wordlength of an FIR filter by utilizing the autocorrelation property of the impulse response is presented. The desired FIR filter is realized as a cascade of a low order all-pole filter modeling the desired FIR filter and another FIR filter representing the modeling error. This technique can also be used simultaneously with other wordlength reduction techniques to enhance the wordlength reduction capability. The resulting filter may be implemented as a parallel combination of a limited number of filters whose coefficient values are

, or

. The coefficient values can also be made equal to powers-of-two.