In this paper, Markov Chain Monte Carlo techniques are applied to feature estimation for automatic speech recognition. By using these methods, it is possible to explore new possibilities in leveraging the autoregressive assumption for noise robust feature extraction. Two minimum mean square error estimators are compared that directly estimate the mean of the feature vectors. The first estimator uses the assumption that the speech is an autoregressive signal, while the second makes no assumptions about the speech spectrum. By creating samples from the posterior distribution, these methods also provide an elegant solution to finding feature variances. These variances can be used to create optimal temporal smoothers of the features as well as input for uncertainty observation decoding. Testing on the Aurora2 database shows that autoregressive modeling provides additional information to improve speech recognition performance. In addition, both smoothing and uncertain observation decoding improve performance in this method.