Improving LPC analysis of speech in additive noise

Linear prediction based speech (LPC) analysis is known to be sensitive to the presence of additive noise. In this paper, we present a noise-compensated method for LPC analysis which ensures good spectral matching between the original speech spectrum and the autoregressive (AR) model spectrum. In this method, the noise periodogram is obtained first by applying a simplified noise power spectral density (PSD) estimator on the calculated noisy periodogram. Then, the effect of noise on the spectral parameters is decreased by gradually subtracting values of the resulting noise autocorrelation coefficients from the coefficients derived from the noisy speech. By taking the absolute value of the estimated reflection coefficients as the decision criterion, we show that this iterative procedure ensures a significant decrease of the degrading effect of noise while the estimated autocorrelation matrix is guaranteed to be positive definite. The method was tested on real speech signals and yielded superior performance when compared to conventional LPC analysis, even in severe noisy conditions.