Uninformed abnormal event detection on audio

Although abnormal events in an audio stream are by their nature hard to define, a continuous monitoring of audio surveillance data can detect crucial information in, e.g., train engines that might require critical maintenance. Our method detects abnormal events without being trained on a certain situation, by building a model of the expected sound environment given a continuously adapting history of past audio material within a limited time interval. We evaluate the precision of this method on recordings from train rides.