TAU-SEBin Binaural Sound Events 2021

TAU-SEBin Binaural Sound Events 2021 is a dataset of synthetic binaural audio recordings, which consist of sound events spaced in simulated shoebox rooms. The data is suitable for experiments with several acoustic scene analysis tasks such as sound source localization, sound distance estimation or sound event detection.

Data is created using isolated sound events derived from several datasets: NIGENS [1], DESED [2] and TUT Rare Sound Events 2017 [3], containing 18 total sound classes, namely: alarm, baby, blender, cat, crash, dishes, dog, engine, fire, footsteps, glassbreak, gunshot, knock, phone, piano, scream, speech, water. The data is split into two subsets, one of which (bin_prox_dir) contains up to two overlapping sound events, whereas the other one consists of single sources only (bin_prox_dir_one). Each subset contains 400 audio files, divided into 4 equal splits for fold-wise cross-validation.

The metadata provides the following information:

sound_event_recording - sound event class

start_time, end_time - onset and offset times of the sound events (in seconds)

azi, ele - the azimuth and elevation angle of the sound source (in degrees)

dist - sound source to receiver distance (in metres)

References:

[1] I. Trowitzsch, J. Taghia, Y. Kashef, and K. Obermayer, NIGENS general sound events database. Zenodo, 2019.
[2] N. Turpault, R. Serizel, A. Parag Shah, and J. Salamon, “Sound event detection in domestic environments with weakly labeled data and soundscape synthesis,” in Workshop on Detection and Classification of Acoustic Scenes and Events, 2019.
[3] A. Mesaros, T. Heittola, A. Diment, B. Elizalde, A. Shah, E. Vincent, B. Raj, and T. 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), 2017, pp. 85–92.