Estimating the energy contour of noise-corrupted speech signals by autocorrelation extrapolation

In processing noise-corrupted speech signals to enhance their intelligibility using a feature-based processing system, it has been determined that an important feature is the energy contour of the speech. That is, to improve intelligibility it is important that the processed signal have an energy contour that matches as closely as possible that of the original uncorrupted speech signal. A method is proposed which explicitly extrapolates the autocorrelation of the noise-corrupted signal to lag zero in order to estimate the noise variance. This noise variance is subtracted from the noisy energy contour to produce an estimate of the original energy contour. This method is shown to outperform conventional methods of noise variance estimation during speech, particularly in the case of high noise levels