Dataset Hub

This dataset comprises audio recordings together with corresponding labels, direction vectors and angle measurements. Labels range from 0 to 96. The training split contains 22,500 samples, totalling 21.5 GB; the download size is approximately 21.5 GB.

The SpatialSounds dataset includes the balanced training and evaluation sets of AudioSet, as well as reverberation data. The dataset provides audio files and associated metadata, supporting mono, binaural, and surround sound formats. In addition, sample code for generating spatial audio and the SpatialSoundQA dataset are provided for training BAT models.