Choice of a finite-accuracy echo-cancelling algorithm

A classical gradient echo-cancelling algorithm and an algorithm which uses only the sign of the error are compared quantitatively in terms of residual-echo (RE) level, complexity of finite-accuracy computations, and speed of convergence. Simulation results are presented graphically and discussed. The sign algorithm is shown to have lower total complexity than the gradient algorithm and to exhibit similar RE and binary-word-length behavior when controlled forced noise is added. If its lower convergence speed is corrected for by introducing a learning phase (without forced noise) and a variable-incrementation step (during the permanent phase), the lower complexity of the sign algorithm makes it superior.