In this paper, the autoregressive (AR) model driven by a white noise with a time-varying variance is proposed for speech signal modeling. An iterative parameter estimation method is then presented, based on the analysis of the short segment (2~4ms) of Linear Prediction (LP) residual. This approach is then used in the framework of speech enhancement based on Kalman filter. A comparative study with existing methods is proposed and shows that the adopted model provides a higher improvement of Signal to Noise Ratio (SNR) than the ordinary AR model.