New directions to autocorrelation function estimation

Abstract

Motivated by an application in speech recognition, we have been searching for a "better" estimate of the autocorrelation function. We give a comparison between the commonly used estimate and our new approach. It is shown that this new technique succeeds in making a compromise between two requirements for real-time signal processing - short length of the analyzed signal frame, and the quality of the estimate - that are often used for description of the pseudo-stationary parts of the analyzed signal.