Avoiding slow band-edge convergence in subband echo cancelers

Subband echo cancelers decompose their input signals into frequency bands and then do echo canceling on a per-band basis. Such structures have both computational and algorithmic advantages over conventional full-band structures. However, many implementations have been found to suffer from slow convergence at band edges, that is, at frequencies carried roughly equally by two adjacent subbands. A poorly conditioned correlation matrix with eigenvalues approaching zero is generally cited as the culprit. One possible solution to the poor band-edge convergence problem is a scheme dubbed postfiltering by Morgan (1995) and De Leon (1995). Under postfiltering, the slowly converging eigenmodes continue to be excited, but their energy is removed from the final output. An alternative solution developed here makes the subband analysis filter for the reference signal slightly broader in spectrum than the analysis filter for the echo. With unequal analysis filtering, it is possible to avoid eliciting the bad eigenmodes in the first place. There are important practical advantages for unequal analysis filtering. Extreme care must be exercised in filter design. However, given a proper understanding of requirements, it is possible to synthesize reasonable-length FIR filters with entirely satisfactory performance