Abstarct: One of the most important issues in the field of speech recognition is the effect of noise on speech signal and degradation of the speech recognition rate. Therefore, robust speech recognition is an active branch of researches. In recent years many researches in this field have been done and different methods are proposed. A group of existing methods in the field of robust speech recognition is the extraction of speech features which are robust against noise. The aim of this thesis is to propose some new methods to extract robust features for speech recognition when the noise is additive to speech signal. In this thesis the autocorrelation domain processing has been used for robust feature extraction. For this purpose, after studying the most important components of speech recognition systems and explaining some of the works that done in the field of robust feature extraction in autocorrelation domain, the proposed methods are presented. The most important methods that used autocorrelation domain for robust feature extraction are RAS, DAS, AMFCC and PAC. In this thesis some new ideas for improving PAC and AMFCC methods baxsed on using differential power spectrum and appropriate windowing has been proposed. The results obtained from the implementation of the proposed methods on the TIMIT databaxse represents improvement in the continuous speech recognition rate compared to the previews methods.