EMOPIA: Multi-Modal Pop Piano Clips with Emotion Labels

EMOPIA is a dataset of 1,087 pop piano music clips from 387 songs, annotated with clip-level emotion labels by four dedicated annotators. It was created by researchers including Hsiao-Tzu Hung from Academia Sinica and presented at ISMIR in 2021. The dataset includes multi-modal data in audio and MIDI formats.

Use Cases

• Train emotion recognition models based on audio and MIDI features of pop piano music.
• Develop emotion-based music generation systems using the labeled MIDI clips.
• Study the relationship between musical features and perceived emotion in pop piano genres.
• Create train/test splits for music information retrieval tasks while avoiding song overlap.

Strengths

• Contains 1,087 music clips with manually annotated emotion labels.
• Provides multi-modal data (audio and MIDI) for each clip.
• Includes metadata and scripts to create train/val/test splits to prevent data leakage from the same song.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for metadata files is unknown, which may limit suitability assessment.
• Last update date is unknown; freshness unverified.

Provenance

Source

Academia Sinica

Collection Method

Clips were crawled from YouTube and transcribed using GiantMIDI. Emotion labels were annotated by four dedicated annotators.