A new algorithm, for linear system parametric identification, is proposed which is statistically consistent if an additive noise is present in the input-output data sequences. Theoretical results are in good agreement with several experimental applications carried out with different levels of noise.