Spectral deconvolution of a multistage nonstationary process based on instantaneous maximum entropy estimation

A method to estimate the spectrum of a multistage nonstationary process is developed. A cascaded autoregressive process is adopted as the model of a process for an observed signal. Each process is supposed to be nonstationary. An instantaneous maximum entropy method based on an instantaneously defined evaluation function and a time-variant lattice filter is proposed. A novel approach is proposed for the problem of process separation based on the nonstationarity of each process using a priori knowledge. The algorithm is proved to be efficient through evaluations using an artificially synthesized test signal and a speech signal