This document presents a new method for creating robust speech features based on linear predictive coding (LPC) for noisy speech recognition. The method applies a weighted arcsine transform to the autocorrelation sequence (ACS) of each speech frame. This transform uses an SNR-dependent smoothing factor to more heavily smooth segments with lower SNR. It also weights each ACS component by the inverse of the average magnitude difference function (AMDF) to emphasize spectral peaks. Experimental results on Mandarin digit recognition show the new LPC features are more noise robust than conventional LPC features over a wide range of SNRs.