Widespread availability and use of music have made automated audio genre classification an important field of research. Thanks to feature extraction systems, not only music data, but also features for them have become readily available. However, handlabeling of a large amount of music data is time consuming. In this study, we introduce a semi-supervised random feature ensemble method for audio classification which uses labeled and unlabeled data together for better genre classification. In order to have diverse subsets of features which are both relevant and non-redundant within themselves, we introduce the Prob-mRMR (Probabilistic minimum Redundancy Maximum Relevance) feature selection algorithm. ProbmRMR is based on mRMR of Ding and Peng 2003 and it selects the features probabilistically according to relevance and redundancy measures. Experimental results show that ensembles of classifiers using Prob-mRMR feature subsets outperform both Co-training and RASCO (Random Subspace Method for Co-training, Wang 2008) which uses random feature subsets.