This paper addresses the effects of the quantization of an audio signal on the Least-Squares (LS) estimate of its autoregressive (AR) model. First, three topics are reviewed: the statistical description of the quantization error in terms of the number of bits used in fixed-point representation for a signal; the LS estimation of the AR model for a signal; and the relation between Minimum Mean-Square Error (MMSE) solutions for the AR model obtained from noisy and noiseless signals. The sensitivity of the associated generator filter poles localization (expressed by magnitudes and phases) to the deviation of the model parameters is examined. Through the interconnection of these aspects, the deviation of the model coefficients is described in terms of the number of bits used to represent the signal to be modeled, which allows for model correction. Conclusions about peculiarities of the pole deviation of the generator filter are drawn.
Autoregressive models; interpolation algorithms; least-squares estimation; quantization noise; sensitivity analysis
