We present a new method of analysis of short data records that will enable us to process nonstationary signals like speech. The method uses standard autocorrelation-based linear prediction (LP) analysis in two phases. In the first phase a low order LP analysis is used to pick up the gross features of the spectral envelope. The LP residual from the first phase is used to modify the LP spectrum in the second phase to capture the details over a short segment of analysis frame. By this method we try to reduce the finite window effects for short time segment analysis. We make use of the fact that the correlation within the samples of the LP residual is less compared to that of the actual signal. Hence effects of short window size are smaller in the LP residual than in the case of the actual signal.