J. Lee, Y.-H. Chen, and Y.-W. Chiu (Taiwan)
Frequency domain adaptive filters, convergence bounds,convergence properties, BLMS, and GMDF.
Frequency domain adaptive filters have gained much attention in applications require large filter coefficients such as acoustic echo cancellation. However, just like any other block processing LMS algorithms, the step-size of frequency LMS has to be limited between 0 and 2 to assure the convergence. This limitation reduces the usage of frequency LMS. In the paper, we revisit the GMDF algorithm, investigate its convergence properties, and provide extensive simulation results to show that for white input signals, NBLMS and GMDF all run well even though the step-size utilized are much bigger than that currently available in the literature. For correlated input signals, frequency normalized GMDF outperforms its counterparts, and still runs very well with big step-size.
Important Links:
Go Back
