RealDESED: A Real-World Domestic Sound Event Detection Benchmark

This dataset contains 5,710 audio snippets recorded in domestic environments, along with 64,430 annotated regions, each consisting of an onset, an offset, and a class label (one of 15 target classes). Additionally, each audio file includes the following metadata provided by person recording the sound: A short textual description of the audible activity A list of audible target-classes A list of sound event classes outside the 15 predefined ones that are audible in the recording The recording location (room) The recording device used Dataset Structure The dataset is divided into three predefined splits: training, validation, and test, containing 3,704, 999, and 1,007 audio recordings respectively. If you use this dataset, we recommend using this pre-defined split to avoid information leakage and allow comparison with other methods. Each split comes as a ZIP-package which has the following structure: <split>.zip ├── audio/ │ ├── 000001.wav │ ├── 000002.wav │ └── ... ├── annotations.csv ├── annotations_raw.csv └── metadata.csv The audio folder contains the waveforms. metadata.csv contains information about individual audio recordings such as the user-specified attributes mentioned above. Annotations in the test and validation sets were manually reviewed and corrected by human reviewers prior to aggregation into the per-file ground truth. The resulting majority-vote-aggreagated annotations are available in annotations.csv. The same applies to approximately 30% of the development set. The raw annotations and reviews before aggregation are available in annotations_raw.csv. The instructions provided to the data collectors and annotators are detailed in Data_Collection_Guidelines.pdf and Data_Annotation_Guidelines.pdf. License The usage terms for the audio recordings in audio folder and their corresponding metadata are specified in the accompanying metadata.csv file (see the license column). The license is either CC0 or CC BY; in the latter case the creator's name is also provided in the license column. The remaining information in metadata.csv and the annotations (provided in annotations_raw.csv and annotations.csv) are licensed under CC BY 4.0, which permits reuse provided that appropriate attribution is given. Accordingly, if you use this dataset, cite this resource and the associated paper.