Improving the Quality of Speech Synthesis Based on the Phase Roots of the Predictor Filter System Function

In digital voice communication systems, in order to parametrically identify the speech signal model, in order to reduce the data rate, the parameters of the predictor filter, called linear prediction coefficients (LPC), are calculated. The paper proposes a new method for parametric identification of a speech signal model based on the phase roots (PhR) of the predictor filter system function (PFSF). It is shown that in the case of distorted LPC and PhR, the method of speech synthesis according to the PhR has a better quality compared to the method of speech synthesis according to the LPC. Thus, with relative parameter errors from 0.01 to 0.1, the mean square error (MSE) of speech synthesis according to the PhR is 8.3 times less than the MSE of speech synthesis according to the LPC.