Realization of the NLMS based transversal adaptive filter using block floating point arithmetic

This paper presents a novel scheme to implement the normalized LMS algorithm in block floating point (BFP) format which permits processing of data over a wide dynamic range at a processor cost marginally higher than that of a fixed point processor. Appropriate BFP formats for both the data and the filter coefficients have been adopted and adjustments made in filtering as well as weight updatation processes so as to sustain the adopted formats and to prevent overflow in these two operations jointly. This is achieved by restricting the step size control parameter available in the NLMS algorithm to lie within an upper bound, which is less than the upper bound for convergence only slightly and thus has marginal effect on convergence speed.