README.md

---
title: Song Prep
emoji: 🎵
colorFrom: purple
colorTo: gray
sdk: docker
app_port: 7860
models:
- tencent/SongPrep-7B
---

# SongPrep
<p align="center"><img src="img/logo.jpg" width="40%"></p>
<p align="center">
    <a href="https://song-prep.github.io/demo/">Demo</a> &nbsp;|&nbsp; <a href="https://arxiv.org/abs/2509.17404">Paper</a>  &nbsp;|&nbsp; <a href="https://huggingface.co/waytan22/SongPrep-7B">Weight</a>  &nbsp;|&nbsp; <a href="https://huggingface.co/datasets/waytan22/SSLD-200">Dataset</a>
</p>
This repository is the official code repository for SongPrep: A Preprocessing Framework and End-to-end Model for Full-song Structure Parsing and Lyrics Transcription. SongPrep is able to analyze the structure and lyrics of entire songs and provide precise timestamps without the need for additional source separation. In this repository, we provide the SongPrep model, inference scripts, and checkpoints trained on the Million Song Dataset that support both Chinese and English.


## Evaluation
Results are reported in Diarization Error Rate (DER) for structure parsing and Word Error Rate (WER) for lyrics transcription.
| Model            | #Params | WER | DER |
|:----------------:|:-------:|:--------:|:--------:|
| SongPrep   | 7B | **23.5%** | **18.2%** |
| Gemini-2.5   | - | 29.2% | 94.6% |
| Seed-ASR         | 12B+ | 104.1% | - |
| Qwen3-ASR       | - | 33.3% | - |
| Qwen-Audio    | 8.4B | 232.7% | - |


## Installation

### Start from scratch

You can install the necessary dependencies using the `requirements.txt` file with Python>=3.8.12 and CUDA>=11.8:
```bash
pip install -r requirements.txt
```

## Usage
To ensure the model runs correctly, please **download the weight** from the original source at [Hugging Face](https://huggingface.co/waytan22/SongPrep-7B), and save it into **root directory** of the project.

Once everything is set up, you can run the inference script using the following command:

### With transformers
```bash
python3 run.py -i your_wav_path
```
The complete output may look like:
```bash
[structure][start:end]lyric ; [structure][start:end]lyric ; [structure][start:end]lyric
```
- The song is divided into segments by ';'.
- The **structure** is the label from structure analysis for the segment.
- The **start** and **end** are the segment’s start and end times.
- The lyric is the recognized lyrics, where sentences separated by '.'.


### With vllm
```bash
python3 run.py -i your_wav_dir
```

The complete output may look like:
```bash
=====wav_name=====
[structure][start:end]lyric ; [structure][start:end]lyric ; [structure][start:end]lyric

=====wav_name=====
[structure][start:end]lyric ; [structure][start:end]lyric ; [structure][start:end]lyric

=====wav_name=====
[structure][start:end]lyric ; [structure][start:end]lyric ; [structure][start:end]lyric
```

## Citation
```
@misc{tan2025songpreppreprocessingframeworkendtoend,
      title={SongPrep: A Preprocessing Framework and End-to-end Model for Full-song Structure Parsing and Lyrics Transcription}, 
      author={Wei Tan and Shun Lei and Huaicheng Zhang and Guangzheng Li and Yixuan Zhang and Hangting Chen and Jianwei Yu and Rongzhi Gu and Dong Yu},
      year={2025},
      eprint={2509.17404},
      archivePrefix={arXiv},
      primaryClass={eess.AS},
      url={https://arxiv.org/abs/2509.17404}, 
}
```

## License
The code and weights in this repository is released in the [LICENSE](LICENSE)  file.