Windowing in Linear Prediction Analysis of Voiced Speech

Abstract

The autocorrelation and covariance methods of linear prediction are used for pitch-synchronous and pitch-asynchronous analyses of voiced-speech segments and the effect of weighting the data prior to its spectral analysis by a tapered window function is studied. It is shown that for both pitch-synchronous and pitch-asynchronous analyses, the performance of the autocorrelation method improves considerably by the introduction of such a window function. In case of the covariance method, no such windowing of data is required. Furthermore, it is shown that for pitch-synchronous analysis, the performance of the autocorrelation method is not much worse than that of the covariance method. Hence, the use of autocorrelation method is recommended over the covariance method if stability of the linear prediction filter is to be ensured.