Mar2Ding/songcompose_data

[ACL 2025] SongCompose Dataset

This repository hosts the official dataset used in SongComposer, a system designed for aligning lyrics and melody for LLMs-based vocal composition.

🌟 Overview

The dataset includes three types of aligned resources, grouped by language (English and Chinese):

• lyric: Unpaired lyrics in English and Chinese
• melody: Unpaired melodies (note sequences and durations)
• pair: Aligned lyric-melody pairs with note durations, rest durations, structures, and tempo (bpm)

Each top-level key in the JSON corresponds to a category (lyric, melody, pair), and each of them has nested keys en and cn for language split.

📂 Structure

Example format:

{
  "pair": {
    "en": [
      {
        "lyrics": "...",
        "notes": [...],
        "notes_duration": [...],
        "rest_duration": [...],
        "bpm": 120
      },
      ...
    ],
    "cn": [ ... ]
  },
  "melody": {
    "all": [ ... ],
  },
  "lyric": {
    "en": [ ... ],
    "cn": [ ... ]
  }
}

📖 Citation

If you use this dataset in your work, please cite our paper:

@misc{ding2024songcomposer,
      title={SongComposer: A Large Language Model for Lyric and Melody Generation in Song Composition}, 
      author={Shuangrui Ding and Zihan Liu and Xiaoyi Dong and Pan Zhang and Rui Qian and Junhao Huang and Conghui He and Dahua Lin and Jiaqi Wang},
      year={2024},
      eprint={2402.17645},
      archivePrefix={arXiv},
      primaryClass={cs.SD}
}