Time signal filtering by relative neighborhood graph localized linear approximation

A time signal filtering algorithm based on the relative neighborhood graph (RNG) used for localization of linear filters is proposed. The filter is constructed from a training signal during two stages. During the first stage an RNG is constructed. During the second stage, localized linear filters are associated each RNG node and adapted to the training signal. The filtering of a test signal is then carried out by inserting the test signal vectors in the RNG followed by the determination of the filter output as a function of the linear filters or the RNG nodes to which the vectors are associated. Training examples are given on a segment of a speech signal and a signal with burst structure generated from a bilinear Subba Rao model