AILabs.tw Pop1K7

AILabs.tw Pop1K7 is a dataset comprising 1747 transcribed piano performances of Western, Japanese and Korean pop songs. It was compiled in Compound Word Transformer to support research on composing expressive pop piano music at full-song length. The songs average about 4 minutes in duration, totaling 108 hours of music. All pieces are in a 4/4 time signature (four beats per bar). Each song (an audio) was converted into a symbolic sequence following specific instructions. 

The first file, Pop1K7.zip, contains these processed MIDI files at each step, along with their REMI and CP representations for unconditional generation tasks. The second file, Pop1K7-emo.zip, includes REMI and functional representations specifically designed for emotion-driven conditional generation tasks, as well as detected key signatures. For general conditional generation tasks, simply remove the <Emotion_*> token in each .pkl file. Please refer to the paper Compound Word Transformer for definitions of unconditional and conditional generation.

Citation

@inproceedings{compoundword2021,
    author = {Wen-Yi Hsiao and Jen-Yu Liu and Yin-Cheng Yeh and Yi-Hsuan Yang},
    title = {{Compound Word Transformer}: Learning to Compose Full-Song Music over Dynamic Directed Hypergraphs},
    booktitle = {Thirty-Fifth {AAAI} Conference on Artificial Intelligence, {AAAI}},
    year = {2021}
}

@inproceedings{emodisentanger2024,
  author = {Jingyue Huang and Ke Chen and Yi-Hsuan Yang},
  title = {Emotion-driven Piano Music Generation via Two-stage Disentanglement and Functional Representation},
  booktitle = {Proceedings of the International Society for Music Information Retrieval Conference, {ISMIR}},
  year = {2024}
}