Autoregressive nonsteady methods for the analysis and synthesis of the human voice

The autoregressive steady model and its extension to the nonsteady case are described. Starting from Grenier's Theorem it is shown that a generalization of the previous model may describe a nonstationary stochastic process. The hypothesis for the implementation of the nonsteady filter is that the model coefficients could be estimated as weighted linear combinations of a small number of known functions. A Levinson algorithm capable of solving the matrix problem is presented. Sampling interval can be extended to 100 msec.