Exploiting recursive parameter trajectories in speech analysis

Abstract

A novel approach to extracting voicing, pitch, and pitch pulse location information by processing the trajectories produced by weighted recursive least squares algorithms is presented. Traditional approaches to extracting this information usually involve performing an LPC analysis, calculating the residual using the LPC parameters, and then processing the residual using various methods. In this work it is shown that if an RLS algorithm is used for the LPC analysis, then extracting the pitch information from the parameter trajectories produced by the algorithm is an accurate alternative to residual-based pitch estimation. This approach also produces accurate LPC parameter estimates by selecting the RLS coefficients at the point just prior to the pitch pulse indicator. >