ON THE ESTIMATION ERROR ASSOCIATED WITH THE AUTOCORRELATION METHOD LPC ESTIMATE

摘要

Linear predictive coding (LPC) has become the predominant methodology for computer speech analysis.This paper evaluates the performance of the autocorrelation method of LPC analysis and the effects which certain attributes (e.g. frame length, etc.) have on this performance. Theoretical lower bounds on the mean-square estimation error are derived for LPC coefficients and for the vocal-tract filter gain. It is shown that a significant amount of the variability of measured parameters from real speech is due to estimation uncertainty. Finally, a speaker recognition experiment is conducted to demonstrate that better recognition accuracy can be achieved by using longer frame lengths.