Studies of identification algorithms for autoregressive models when the observations are noisy

The effects of noise on an autoregressive signal are studied theoretically and approaches used to improve autoregressive parameter identification in the presence of noise are summarized. Identification methods which do not seem to have been used in this context are discussed including the subtraction of correlation in order to simultaneously estimate the noise and signal parameters, and the application of a principle for identifying a linear system for its free (transitory) response. Monte Carlo type identification experiments were conducted on synthetic autoregressive signals, with different signal to noise relationships. A series of curves shows the mean performances of the methods as a function of signal to noise relationship and as a function of other significant parameters. The performances were evaluated in speech analysis/synthesis.