| # SongFormBench 🏆 |
|
|
| [[English](README.md) | 中文] |
|
|
| **一个高质量的音乐结构分析基准** |
|
|
| <div align="center"> |
|
|
|  |
|  |
| [](https://arxiv.org/abs/2510.02797) |
| [](https://github.com/ASLP-lab/SongFormer) |
| [](https://huggingface.co/spaces/ASLP-lab/SongFormer) |
| [](https://huggingface.co/ASLP-lab/SongFormer) |
| [](https://huggingface.co/datasets/ASLP-lab/SongFormDB) |
| [](https://huggingface.co/datasets/ASLP-lab/SongFormBench) |
| [](https://discord.gg/p5uBryC4Zs) |
| [](http://www.npu-aslp.org/) |
|
|
| </div> |
|
|
| <div align="center"> |
| <h3> |
| Chunbo Hao<sup>1*</sup>, Ruibin Yuan<sup>2,6*</sup>, Jixun Yao<sup>1</sup>, Qixin Deng<sup>3,6</sup>,<br>Xinyi Bai<sup>4,6</sup>, Yanbo Wang<sup>5</sup>, Wei Xue<sup>2</sup>, Lei Xie<sup>1†</sup> |
| </h3> |
| |
|
|
| <p> |
| <sup>*</sup>Equal contribution <sup>†</sup>Corresponding author |
| </p> |
| <p> |
| <sup>1</sup>Audio, Speech and Language Processing Group (ASLP@NPU),<br>School of Computer Science, Northwestern Polytechnical University<br> |
| <sup>2</sup>Hong Kong University of Science and Technology<br> |
| <sup>3</sup>Northwestern University<br> |
| <sup>4</sup>Cornell University<br> |
| <sup>5</sup>University of New South Wales<br> |
| <sup>6</sup>Multimodal Art Projection (M-A-P) |
| </p> |
| |
| </div> |
|
|
| --- |
|
|
| ## 🌟 什么是 SongFormBench? |
|
|
| SongFormBench 是一个**经过精心整理、由专家标注的基准数据集**,旨在彻底改变音乐结构分析(MSA)的评估方式。我们的数据集为比较 MSA 模型提供了统一标准。 |
|
|
| ### 📊 数据集构成 |
|
|
| - **🎸 SongFormBench-HarmonixSet (BHX)**: 来自 HarmonixSet 的 200 首歌曲 |
| - **🎤 SongFormBench-CN (BC)**: 100 首中文流行歌曲 |
|
|
| **总计:300 首高质量标注歌曲** |
|
|
| --- |
|
|
| ## ✨ 主要亮点 |
|
|
| ### 🎯 **统一评估标准** |
| - 建立了 **标准化基准**,实现 MSA 模型间的公平比较 |
| - 消除了评估协议中的不一致性 |
|
|
| ### 🏷️ **简单标签系统** |
| - 采用广泛使用的7类分类系统(如 [arxiv.org/abs/2205.14700](https://arxiv.org/abs/2205.14700) 中所述) |
| - 保留 **pre-chorus** 段落以增强粒度 |
| - 可轻松转换为 7 类(pre-chorus → verse)以保证兼容性 |
|
|
| ### 👨🔬 **专家验证质量** |
| - 多源验证 |
| - 专家标注员手动校正 |
|
|
| ### 🌏 **多语言覆盖** |
| - **首个中文 MSA 数据集**(100 首歌曲) |
| - 弥补了中文音乐结构分析的空白 |
| - 支持跨语言 MSA 研究 |
|
|
| --- |
|
|
| ## 🚀 快速开始 |
|
|
| ### 快速加载 |
| ```python |
| from datasets import load_dataset |
| |
| # 加载完整基准数据集 |
| dataset = load_dataset("ASLP-lab/SongFormBench") |
| ``` |
|
|
| --- |
|
|
|
|
| ## 🤝 引用 |
|
|
| ```bibtex |
| @misc{hao2026songformerscalingmusicstructure, |
| title={SongFormer: Scaling Music Structure Analysis with Heterogeneous Supervision}, |
| author={Chunbo Hao and Ruibin Yuan and Jixun Yao and Qixin Deng and Xinyi Bai and Yanbo Wang and Wei Xue and Lei Xie}, |
| year={2026}, |
| eprint={2510.02797}, |
| archivePrefix={arXiv}, |
| primaryClass={eess.AS}, |
| url={https://arxiv.org/abs/2510.02797}, |
| } |
| ``` |
| --- |
|
|
|
|
| ## 🎼 梅尔频谱图细节 |
|
|
| <details> |
| <summary>Click to expand/collapse</summary> |
|
|
| 环境配置可参考 BigVGan 的官方实现。如果音频源失效,可使用以下方法重建音频。 |
|
|
| ### 🎸 SongFormBench-HarmonixSet |
| 使用官方 HarmonixSet 梅尔频谱图。复现方法如下: |
|
|
| ```bash |
| # 克隆 BigVGAN 仓库 |
| git clone https://github.com/NVIDIA/BigVGAN.git |
| |
| # 进入 utils 目录 |
| cd utils/HarmonixSet |
| |
| # 更新 inference_e2e.sh 中的 BIGVGAN_REPO_DIR |
| # 运行推理脚本 |
| bash inference_e2e.sh |
| ``` |
|
|
| ### 🎤 SongFormBench-CN |
| 使用 [**bigvgan_v2_44khz_128band_256x**](https://huggingface.co/nvidia/bigvgan_v2_44khz_128band_256x) 重建。 |
|
|
| 您应首先下载 bigvgan_v2_44khz_128band_256x,然后将其项目目录添加到 PYTHONPATH 中,之后即可使用以下代码: |
| ```python |
| # 查看实现 |
| utils/CN/infer.py |
| ``` |
| </details> |
|
|
|
|
| ## 📧 联系方式 |
|
|
| 如有问题、反馈或合作机会,请访问我们的 [GitHub 仓库](https://github.com/ASLP-lab/SongFormer) 或提交问题。 |