SongFormBench

# SongFormBench 🏆

[English ｜ [中文](README_ZH.md)]

**A High-Quality Benchmark for Music Structure Analysis**

<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^1*, Ruibin Yuan^2,5*, Jixun Yao¹, Qixin Deng^3,5,<br>Xinyi Bai^4,5, Wei Xue², Lei Xie^1†
</h3>

<p>
^*Equal contribution &nbsp;&nbsp; ^†Corresponding author
</p>

<p>
¹Audio, Speech and Language Processing Group (ASLP@NPU),<br>Northwestern Polytechnical University<br>
²Hong Kong University of Science and Technology<br>
³Northwestern University<br>
⁴Cornell University<br>
⁵Multimodal Art Projection (M-A-P)
</p>
</div>

---

## 🌟 What is SongFormBench?

SongFormBench is a **carefully curated, expert-annotated benchmark** designed to revolutionize music structure analysis (MSA) evaluation. Our dataset provides a unified standard for comparing MSA models.

### 📊 Dataset Composition

- **🎸 SongFormBench-HarmonixSet (BHX)**: 200 songs from HarmonixSet
- **🎤 SongFormBench-CN (BC)**: 100 Chinese popular songs

**Total: 300 high-quality annotated songs**

---

## ✨ Key Highlights

### 🎯 **Unified Evaluation Standard**
- Establishes a **standardized benchmark** for fair comparison across MSA models
- Eliminates inconsistencies in evaluation protocols

### 🏷️ **Simple Label System**
- Adopts the widely used 7-class classification system (as described in [arxiv.org/abs/2205.14700](https://arxiv.org/abs/2205.14700)
)
- Preserves **pre-chorus** segments for enhanced granularity
- Easy conversion to 7-class (pre-chorus → verse) for compatibility

### 👨‍🔬 **Expert-Verified Quality**
- Multi-source validation
- Manual corrections by expert annotators

### 🌏 **Multilingual Coverage**
- **First Chinese MSA dataset** (100 songs)
- Bridges the gap in Chinese music structure analysis
- Enables cross-lingual MSA research

---

## 🚀 Getting Started

### Quick Load
```python
from datasets import load_dataset

# Load the complete benchmark
dataset = load_dataset("ASLP-lab/SongFormBench")
```

---

## 📚 Resources & Links

- 📖 Paper: *coming soon*
- 💻 Code: [GitHub Repository](https://github.com/ASLP-lab/SongFormer)
- 🧑‍💻 Model: [SongFormer](https://huggingface.co/ASLP-lab/SongFormer)
- 📂 Dataset: [SongFormDB](https://huggingface.co/datasets/ASLP-lab/SongFormDB)

---

## 🤝 Citation

```bibtex
@misc{hao2025songformer,
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 Wei Xue and Lei Xie},
year = {2025},
eprint = {2510.02797},
archivePrefix = {arXiv},
primaryClass = {eess.AS},
url = {https://arxiv.org/abs/2510.02797}
}
```

---

## 🎼 Mel Spectrogram Details

<details>
<summary>Click to expand/collapse</summary>

Environment configuration can refer to the official implementation of BigVGan. If the audio source becomes invalid, you can reconstruct the audio using the following method.

### 🎸 SongFormBench-HarmonixSet
Uses official HarmonixSet mel spectrograms. To reproduce:

```bash
# Clone BigVGAN repository
git clone https://github.com/NVIDIA/BigVGAN.git

# Navigate to utils
cd utils/HarmonixSet

# Update BIGVGAN_REPO_DIR in inference_e2e.sh
# Run the inference script
bash inference_e2e.sh
```

### 🎤 SongFormBench-CN
Reproduce using [**bigvgan_v2_44khz_128band_256x**](https://huggingface.co/nvidia/bigvgan_v2_44khz_128band_256x)

You should first download bigvgan_v2_44khz_128band_256x, then add its project directory to your PYTHONPATH, after which you can use the code below:

```python
# See implementation
utils/CN/infer.py
```
</details>

---

## 📧 Contact

For questions, issues, or collaboration opportunities, please visit our [GitHub repository](https://github.com/ASLP-lab/SongFormer) or open an issue.

# SongFormBench 🏆

[English ｜ [中文](README_ZH.md)]

**高质量音乐结构分析（Music Structure Analysis, MSA）基准数据集**

<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>
郝春博^1*, 袁瑞彬^2,5*, 姚继勋¹, 邓启昕^3,5,<br/>
白欣怡^4,5, 薛巍², 谢磊^1†
</h3>

<p>
^*共同第一作者 &nbsp;&nbsp; ^†通讯作者
</p>

<p>
¹音频、语音与语言处理课题组（ASLP@NPU）,<br/>西北工业大学<br/>
²香港科技大学<br/>
³美国西北大学<br/>
⁴康奈尔大学<br/>
⁵多模态艺术投影（M-A-P）
</p>
</div>

---

## 🌟 什么是SongFormBench？

SongFormBench是一个**经过精心筛选、专家标注的基准数据集**，旨在革新音乐结构分析（MSA）的评估范式。本数据集为音乐结构分析模型的横向对比提供了统一标准。

### 📊 数据集构成

- **🎸 SongFormBench-HarmonixSet（BHX）**：取自HarmonixSet的200首歌曲
- **🎤 SongFormBench-CN（BC）**：100首中文流行歌曲

**总计：300首经过高质量标注的歌曲**

---

## ✨ 核心亮点

### 🎯 **统一评估标准**
- 建立了**标准化基准数据集**，实现音乐结构分析模型间的公平对比
- 消除了评估流程中的不一致性

### 🏷️ **简洁标注体系**
- 采用学界广泛使用的7分类系统（详见[arxiv.org/abs/2205.14700](https://arxiv.org/abs/2205.14700)）
- 保留**预副歌段（pre-chorus）**以提升标注粒度
- 可轻松转换为7分类体系（预副歌段（pre-chorus）→ 主歌（verse））以适配不同模型需求

### 👨‍🔬 **专家验证的高质量标注**
- 采用多源验证机制
- 由专业标注人员进行手动修正

### 🌏 **多语言覆盖**
- 推出**首个中文音乐结构分析（MSA）数据集**（100首歌曲）
- 填补了中文音乐结构分析研究的空白
- 支持跨语言音乐结构分析研究

---

## 🚀 快速上手

### 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")

---

## 📚 资源与链接

- 📖 论文：*即将上线*
- 💻 代码：[GitHub仓库](https://github.com/ASLP-lab/SongFormer)
- 🧑‍💻 模型：[SongFormer](https://huggingface.co/ASLP-lab/SongFormer)
- 📂 数据集：[SongFormDB](https://huggingface.co/datasets/ASLP-lab/SongFormDB)

---

## 🤝 引用格式

bibtex
@misc{hao2025songformer,
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 Wei Xue and Lei Xie},
year = {2025},
eprint = {2510.02797},
archivePrefix = {arXiv},
primaryClass = {eess.AS},
url = {https://arxiv.org/abs/2510.02797}
}

---

## 🎼 梅尔频谱图细节

<details>
<summary>点击展开/折叠</summary>

环境配置可参考BigVGAN的官方实现。若音频源失效，可通过以下方法重构音频。

### 🎸 SongFormBench-HarmonixSet
使用官方HarmonixSet梅尔频谱图。复现步骤如下：

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变量
# 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)进行复现。

你需先下载bigvgan_v2_44khz_128band_256x，将其项目目录添加至PYTHONPATH环境变量，随后可使用以下代码：

python
# 详见实现代码：utils/CN/infer.py

</details>

---

## 📧 联系方式

如有疑问、问题或合作意向，请访问我们的[GitHub仓库](https://github.com/ASLP-lab/SongFormer)或提交Issue。