1984 年 40 巻 6 号 p. 397-406
An autoregressive and moving-average (ARMA) analysis of speech is described in this paper. For the estimation of ARMA parameters, both input and output of the model should be measured. But the analysis of speech can only observe speech; an output of the model, and the process of parameter estimation turns out nonlinear. We propose the method which estimates ARMA parameters and inputs simultaneously. The nonlinear problem in speech analysis can be reduced into two linear problems by this method. But the problems on both input and order estimation of the model may occur in this method. Some fundamentals of the input estimation are discussed. Two independent methods of estimations, the estimations based on pulse series input and Gaussian process input, are proved to be useful for speech analysis. Especially, it is shown that the accurate ARMA parameters can be estimated by using an assumption of pulse series input. On the order of the estimation model, the method is presented which analyzes the speech by a high-order model. The optimal model can be derived from this high-order model.