This paper investigates a noise robust approach to automatic speech recognition based on a mixture of Bayesian joint factor analyzers. In this approach, noisy features are modeled by two joint groups of factors accounting for speaker and noise variabilities which are estimated by clean and noisy speech respectively. The factors form an overcomplete dictionary with a redundant representation. Automatic relevance determination (ARD) is carried out by the relevance vector machine (RVM) where sparsity-promoting priors are applied on two factor loading matrices. Experiments on large vocabulary continuous speech recognition (LVCSR) tasks show good improvements by this approach. Copyright © 2013 ISCA.