README.md

metadata

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

SongPrep

Demo  |  Paper  |  Weight  |  Dataset

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:

pip install -r requirements.txt

Usage

To ensure the model runs correctly, please download the weight from the original source at Hugging Face, 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

python3 run.py -i your_wav_path

The complete output may look like:

[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

python3 run.py -i your_wav_dir

The complete output may look like:

=====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 file.