Comparison of ARMA modeling methods in the time domain

Several ARMA (autoregressive moving average) modeling methods are compared for their ability to model the time response of systems characterized by exponentially damped resonant frequencies. Results are discussed for known ARMA systems with added noise and samples of recorded short-duration acoustic data. It is demonstrated that there are currently available techniques which, when properly applied, are highly effective at modeling even highly structured real-world acoustic signals. In particular, iterative prefiltering was effective at modeling signals with unknown model order, unknown initial conditions, unknown time shifts, and noise present. Effective initial estimates were achieved using overdetermined modified Yule-Walker equations in conjunction with singular value decomposition to estimate model poles and Shank´s method to estimate model zeros