A new filter bank design for split-spectrum algorithm

The most important aspect when designing Split-spectrum algorithms is the filter bank design. There are two main trends commonly followed. One is based on constant bandwidth Gaussian filters equally spaced in frequency and the other one uses wavelets as a multiresolution time-scale method. In this paper we present an alternative that combines both techniques taking advantage from the best of each. It is based in the use of variable bandwidth filters but in this case filters are equally spaced in frequency and are energy gain equalized so that all bands have the same contribution from the possible echo defect. With this new design we improve the insensitivity of the echo defect to the tuning frequency. Improvements in the signal to noise ratio gain after recombination are reported in this contribution. We show the results obtained after simulations using simple stationary models for the grain noise with a single defect. SNR enhancement factor was selected as the figure of merit to make the comparisons among the different methods. These results are contrasted with real processed signal obtained in laboratory from pieces of aluminum alloy, showing the interest of the proposed new filter bank structure and recombination methods.