This study presents a novel approach to unsupervised pronunciation grammar generation for non-native speech recognition. Unsupervised pronunciation grammar generation includes pronunciation variation graph construction, stochastic Markov search and grammar selection. Context-dependent relation and phone broad class information are used for variation graph construction. Confidence measure and co-occurrence frequency are used to select the variants of pronunciation grammar for non-native speech modeling. Experiments show that unsupervised pronunciation grammar generation is suitable for the improvement of non-native speech recognition.