Cancelation of acoustic echoes with exponentially weighted step-size and fixed point arithmetic

Cancelation of acoustic echoes has been proven to be a challenging application for adaptive filters with a large number of coefficients. The typical shape of a room impulse response can be used to speed up convergence of the adaptive filter by weighting the adaptation step-size parameter exponentially. In this contribution we show that the exponential step-size NLMS algorithm loses performance when implemented in fixed point arithmetic. In order to improve its behavior scaling of the filter coefficients can increase the precision of most filter coefficients, so that the degradations caused by fixed point arithmetic in conjunction with the exponentially weighted step-size parameter can be reduced. We also show that filter scaling and exponentially weighted step-size parameter can be easily combined with low overhead, since both algorithms are closely related.