README.md

metadata

pretty_name: SongFormBench
tags:
  - MSA
  - Benchmark
license: cc-by-4.0
language:
  - en
  - zh

SongFormBench 🏆

[English ｜ 中文]

A High-Quality Benchmark for Music Structure Analysis

Chunbo Hao^1*, Ruibin Yuan^2,6*, Jixun Yao¹, Qixin Deng^3,6,
Xinyi Bai^4,6, Yanbo Wang⁵, Wei Xue², Lei Xie^1†

^*Equal contribution    ^†Corresponding author

¹Audio, Speech and Language Processing Group (ASLP@NPU),
School of Computer Science, Northwestern Polytechnical University
²Hong Kong University of Science and Technology
³Northwestern University
⁴Cornell University
⁵University of New South Wales
⁶Multimodal Art Projection (M-A-P)

---

🌟 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 )
• 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

from datasets import load_dataset

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

---

🤝 Citation

@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}, 
}

---

🎼 Mel Spectrogram Details

Click to expand/collapse

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:

# 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

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:

# See implementation
utils/CN/infer.py

---

📧 Contact

For questions, issues, or collaboration opportunities, please visit our GitHub repository or open an issue.