Adaptive filter system having mixed fixed point or floating point and block scale floating point operators

摘要

An adaptive filter is provided which has fixed point or floating point data stored in a data RAM (74) and block scale floating point coefficients stored in a coefficient RAM (84). The data and coefficients are utilized in a filter algorithm which utilizes a multiplier and an accumulator to provide a convolution result. Coefficients are updated by adding the multiplied result of the data RAM value and the error value to the old value of the coefficient. This is done for all the coefficient values in the coefficient RAM. The error value indicates the difference between the filter output and the sampled near-end signal that is the echo. These new coefficients are examined and if any have a value above or all have a value below a predetermined threshold, then the mantissas of all the coefficients are shifted and the exponent adjusted in the next filter cycle.