The Use of a Two-Pole Linear Prediction Model in Speech Recognition
John I. Makhoul, Jared J. Wolf · 1973
Abstract : In speech recognition applications, it is often desirable to make a gross characterization of the shape of the spectrum of a particular sound. The autocorrelation method of linear prediction analysis leads to an all-pole approximation to the signal spectrum. Hence an LPC analysis using two poles produces one possible gross characterization. The two poles are computed as the roots of a quadratic equation whose coefficients are the linear prediction parameters, which are simple functions of the autocorrelation coefficients (R sub 0), (R sub 1), and (R sub 2). The poles are either both real or form a conjugate pair in the z plane. This fact, together with the exact positions of the poles, is particularly useful in describing certain gross characteristics of the spectrum. The spectral dynamic range of the two-pole spectrum and the normalized minimum error are suggested as more suitable substitutes for the two- pole bandwidths in interpreting the information supplied by the model for the purpose of spectral characterization.