This paper describes a pole-zero (ARMA) modeling of speech using a recursive-least-squares (RLS) fast transversal filter (FTF) algorithm. This ARMA FTF algorithm can estimate unknown input excitation and the estimated input is used to determine the parameters of the pole-zero model. This algorithm is derived using geometric projections. The geometric projection approach gives insight and useful interpretation of various filters that form the algorithm. We give a performance evaluation of the proposed algorithm by applying to synthetic and natural speech spectral estimations. This algorithm accurately represents spectral peaks and valleys of speech and requires less computations than RLS lattice filters and ARMA FTF algorithm of Ardalan and Faber (1988). Additionally, this algorithm can also be applied to other signal processing areas where the input is unknown.