The data {y(n)} are fit by the ARMA process with possible model errors {epsilon(n)}:A(z) y(n) = C(z) w(n) + epsilon(n) where {w(n), F(n)} is a martingale difference sequence. The recursive ELS algorithm is used to estimate unknown coefficients of A (z) and C(z). It is shown that the convergence rate of the estimate is [GRAPHICS] and hence the estimate is strongly consistent if SIGMA(i=0)n \\epsilon(i)\\2 = o(n).