MusicCaps is an extensive dataset that comprises 5521 music clips, each lasting for 10 seconds. The clips are labeled with a list of aspects and a free-text caption written by musicians, which describes how the music sounds. The aspect list is a set of adjectives that captures the sonic qualities of the music, such as "pop tinny wide hi hats mellow piano melody high pitched female vocal melody sustained pulsating synth lead." The free-text caption provides additional details about the music, including the instruments and mood.
MusicCaps is derived from the AudioSet dataset and is separated into an evaluation and training split. The dataset is licensed under the Creative Commons BY-SA 4.0 license and each clip is labeled with metadata such as YouTube ID, start and end position in the video, and labels from the AudioSet dataset. Moreover, the dataset includes the aspect list, caption author ID (for grouping samples by who wrote them), is balanced subset, and is AudioSet eval split.
The MusicCaps dataset is particularly useful for music description tasks, and its extensive labeling ensures that it is a valuable resource for anyone interested in music analysis and understanding.
