Dataset Card for MusicCaps

Dataset Description

Overview

MusicCaps contains 5,521 music examples, each annotated with an English aspect list and a free‑text caption authored by musicians. An example aspect list might be "pop, tinny wide hi hats, mellow piano melody, high‑pitched female vocal melody, sustained pulsating synth lead", while a caption may consist of several sentences describing the sound, e.g.:

"A low‑sounding male voice is rapping over fast‑paced drums playing a reggaeton beat along with a bass. Something like a guitar is playing the melody. This recording is of poor audio quality. In the background, laughter can be heard. This song may be playing in a bar."

These annotated examples are extracted from the AudioSet dataset’s 10‑second music clips (2,858 from the eval split, 2,663 from the train split).

Usage

The released data are provided as a .csv file containing YouTube video IDs and start/end timestamps. Using the dataset requires downloading the corresponding YouTube videos and clipping them according to the timestamps.

Supported Tasks and Leaderboards

[More information needed]

Language

[More information needed]

Dataset Structure

Data Instances

[More information needed]

Data Fields

• ytid: YouTube ID of the video containing the annotated music segment. The segment can be listened to at https://youtu.be/watch?v={ytid}&start={start_s}.
• start_s: Start time of the music segment within the YouTube video.
• end_s: End time of the music segment (all segments are 10 seconds long).
• audioset_positive_labels: Labels from the AudioSet dataset for this segment.
• aspect_list: List of aspects describing the music.
• caption: Multi‑sentence free‑text caption describing the music.
• author_id: Integer used to group samples by author.
• is_balanced_subset: If true, the row belongs to a genre‑balanced 1 k subset.
• is_audioset_eval: If true, the segment comes from the AudioSet eval split; otherwise from the train split.

Data Splits

[More information needed]

Dataset Creation

Curation Rationale

[More information needed]

Source Data

Initial Data Collection and Normalization

[More information needed]

Who Produced the Source Language?

[More information needed]

Annotation

Annotation Process

[More information needed]

Who Were the Annotators?

[More information needed]

Personal and Sensitive Information

[More information needed]

Considerations for Using the Dataset

Societal Impact

[More information needed]

Discussion of Biases

[More information needed]

Other Known Limitations

[More information needed]

Additional Information

Dataset Curators

The dataset was shared by @googleai.

License

The dataset is released under CC‑BY‑SA‑4.0.

Citation

[More information needed]

Contributions

[More information needed]