Quantization effects in transform domain adaptive filters

A study of finite wordlength effect in transform domain adaptive filters is presented. Both the frequency domain block LMS (FBLMS) algorithm and the transform domain LMS (TDLMS) algorithm (in which any orthonormal transformation may be used for decorrelating the input samples to the filter) are considered. It is found that the FBLMS, although theoretically requires longer wordlength to prevent overflow of calculations and stalling of the algorithm, is more robust to overflow of the intermediate calculations. Simulation results presented show that the wordlength requirement of the two algorithms is roughly the same