In this paper, we adapt Recurrent Neural Networks with Stochastic Layers,which are the state-of-the-art for generating text, music and speech, to theproblem of acoustic novelty detection. By integrating uncertainty into thehidden states, this type of network is able to learn the distribution ofcomplex sequences. Because the learned distribution can be calculatedexplicitly in terms of probability, we can evaluate how likely an observationis then detect low-probability events as novel. The model is robust, highlyunsupervised, end-to-end and requires minimum preprocessing, featureengineering or hyperparameter tuning. An experiment on a benchmark datasetshows that our model outperforms the state-of-the-art acoustic noveltydetectors.