On the use of autocorrelation for pitch extraction: some statistical considerations and their application to the SIFT algorithm

The aim of this paper is to consider the classical pitch determination method based on the short-time autorcorrelation analysis of the speech signal. Two commonly used estimators and the effect of signal windowing are taken into account. It is shown that in both periodic and purely random cases a similar decomposition holds for the estimated autocorrelation. Such a decomposition makes it possible to foresee the relative merits of the estimators considered, at least as far as gross pitch determination errors and voicing-errors are concerned. The expected behaviour is found to be in good agreement with results obtained through applying the SIFT algorithm to real speech.