Datasets

Magnatune audio clips with human tags collected through the TagATune game for music auto-tagging research.

A unified evaluation suite for pretrained music-audio representations across acoustic, performance, score and high-level tasks.

Chord, structure and feature annotations for a sampled set of Billboard-chart songs.

Annotated royalty-free multitrack recordings supporting melody, pitch, instrument and source-separation research.

Pozalabs MIDI dataset of 4,422 professional-writer samples for fine-level dynamics control.

Audio features and metadata for one million contemporary popular-music tracks, without distributed audio.

Dataset of 1,000 Chinese karaoke song clips with separated accompaniment/vocal channels and manual annotations.

A multi-domain benchmark of expert audio understanding and reasoning spanning speech, environmental sound and music.

Expert-created audio questions covering speech, sound, music, mixtures, spatial audio and long-form reasoning.

Multitrack music dataset with a hierarchical stem taxonomy for fine-grained source-separation research.

Human-validated multiple-choice questions for evaluating music understanding and reasoning in audio-language models.

Full-length stereo music mixtures with isolated drums, bass, vocals and other stems for source separation.

Open benchmark and software package for reproducible evaluation of musical-score inpainting systems.

A live pairwise evaluation platform and rolling leaderboard for text-to-music generation systems.

Rolling human preference data from live pairwise comparisons of text-to-music model outputs.