TUT Rare Sound events 2017, development dataset consists of source files for creating mixtures of rare sound events (classes baby cry, gun shot, glass break) with background audio, as well a set of readily generated mixtures and recipes for generating them. The "source" part of the dataset consists of two subsets: background recordings from 15 different acoustic scenes, recordings with the target rare sound events from three classes, accompanied by annotations of their temporal occurrences, a set of meta files providing the cross-validation setup: lists of background and target event recordings split into training and test subsets (called "devtrain" and "devtest", respectively, indicating they are provided as the development dataset, as opposed to the evaluation dataset released separately). The mixture set consists of two subsets (training and testing), each containing ~1500 mixtures (~500 per target class in each subset, with half of the mixtures not containing any target class events). The collection of the background recording data has been financially supported by European Research Council under the European Unions H2020 Framework Programme through ERC Grant Agreement 637422 EVERYSOUND.

{"references": ["Annamaria Mesaros, Toni Heittola, Aleksandr Diment, Benjamin Elizalde, Ankit Shah, Emmanuel Vincent, Bhiksha Raj, and Tuomas Virtanen. DCASE 2017 challenge setup: tasks, datasets and baseline system. In Proceedings of the Detection and Classification of Acoustic Scenes and Events 2017 Workshop (DCASE2017), pp 85\u201392. November 2017."]}

The license terms are specified in the LICENSE.txt file.