Experimental comparison of stationary and nonstationary formulations of linear prediction applied to voiced speech analysis

Abstract

There are basically two formulations of linear prediction: (1) stationary, and (2) nonstationary. This paper presents the theoretical differences between the two formulations and gives the results of experimental comparison of the two formulations when applied to voiced speech analysis. In this experimental study three criteria are used for comparison purposes: (1) total minimum normalized squared error (η) [Makhoul, “Spectral Analysis of Speech by Linear Prediction,” IEEE Trans. Audio Electroacoust. (June 1973)], (2) accuracy in estimating speech spectrum, and (3) accuracy in estimating formant parameters. This paper presents the results of linear prediction pitch synchronous as well as pitch asynchronous analysis of synthetic and natural speech. Influence of analysis segment size and its position on the estimated formant parameters and η have been investigated. Nonstationary formulation performs better in case of pitch synchronous analysis than pitch asynchronous analysis. We have found that, for pitch synchronous analysis, nonstationarity is a better assumption than stationarity, but for pitch asynchronous analysis and large analysis segment size (20–25 msec) the performance of both formulations in representing the speech waveform is practically the same.