Time-varying Burg method for speech analysis

Abstract

The Burg method determines the reflection coefficients of the lattice filter under the assumption of a stationary process. In this contribution, a time-varying Burg method based on least squares estimation is proposed estimating reflection coefficients which are time-varying within the analyzed frames. For that purpose, the trajectory of the reflection coefficients is parameterized by piecewise linear basis function. To yield estimation results of adjacent frames which are continuously connected in terms of the reflection coefficients, the coefficients of the current frame are constrained in a way that the trajectory starts from the final coefficient vector of the last frame. This leads to a block-based estimation procedure which can be performed in an adaptive way since the estimation of a frame depends only on the current frame and on the last frames.