Power spectral density estimation via wavelet decomposition

Abstract

A soft decision algorithm for wavelet decomposition, in which a probability measure is assigned to each frequency band bearing energy, is presented. This soft decision algorithm is used as an approximate estimator of power spectral density. A staircase approximation of power spectral density (PSD) is obtained by plotting the 2m probabilities after an m-stage decomposition. Different wavelet filters are used for estimating the PSD of a speech segment. The type of the wavelet filter used can be selected as a compromise between accuracy and complexity.