On the analysis of autocorrelation function for speech spectra estimation - application for nasality detection

This paper introduces the Autocorrelation Function Modelling method (AFM) for speech analysis. It has been established that the same AR model can be used to represent both the discrete signal and its autocorrelation function (AF) in a predictive scheme. Using this function and not directly the speech signal for parameter estimations numerous advantages : the system memory, as well as the number of samples involved in the analysis can be relatively small. A system order evaluation can be obtained simultmueously. On the other hand neglecting some terms of the AF results in an estimation of the AR model which is independant of zeros. Comparison of spectra obtained by this method, with spectra obtained by classical Linear Prediction is given below for nasal and non-nasal vowels. For non-nasalized sounds the estimation obtained here is identical to that which was obtained by covariance method.