SingVERSE: A Diverse, Real-World Benchmark for Singing Voice Enhancement

SingVERSE is the first real-world benchmark for singing voice enhancement, created to address the critical lack of realistic evaluation data. It provides a foundational benchmark for developing and evaluating singing voice enhancement models.

The dataset consists of 3,929 audio pairs, totaling 18.14 hours. It spans 19 distinct and diverse real-world acoustic scenarios, from reverberant concert halls to noisy roadsides. Each pair contains a noisy recording captured in a real-world environment and its corresponding studio-quality, time-aligned clean vocal reference.

For more details, please refer to our paper: SingVERSE: A Diverse, Real-World Benchmark for Singing Voice Enhancement Project page: https://singverse.github.io

📊 Dataset Structure

The following table shows the statistics of the SingVERSE dataset, including the number of utterances, average duration (in seconds), and total duration (in seconds) for both Pro and Non-Pro devices.

Scenario	Utterances	Avg.(s)	Total(s)
Concert Hall	62/94	9.9/8.9	611/833
Concert	112/335	10.2/11.8	1139/3966
Arcade	84/84	8.2/8.2	689/689
Restaurant	39/39	5.9/5.9	229/229
Rehearsal Room	120/120	8.3/8.3	991/991
Classroom	66/66	6.7/6.7	440/440
Basement	39/39	5.9/5.9	229/229
Roadside	384/368	8.3/8.3	3180/3036
KTV	230/180	7.7/7.7	1767/1399
Piano Room	178/178	8.3/8.3	1470/1470
Passageway	97/108	7.7/8.0	746/860
Staircase	64/65	7.4/7.4	477/481
Meeting Room	22/22	8.7/8.7	191/191
Parking Lot	68/68	8.9/8.9	606/606
In the Car	39/39	5.9/5.9	229/229
Shopping Mall	60/40	7.4/7.5	442/302
Dormitory	0/54	0.0/10.2	0/551
Office	60/60	7.4/7.4	442/442
Park	123/123	7.5/7.5	917/917

Data Instance

Each row in the dataset represents a single data instance, containing metadata and the corresponding clean/noisy audio.

{
    "singer": "chenyixun",
    "song": "burubujian",
    "scenario": "Concert",
    "device": "pro",
    "part": 0,
    "clean_audio": {
        "path": "/path/to/Concert/clean/chenyixun-burubujian-Concert-pro-part_0.wav",
        "array": ...,
        "sampling_rate": 44100
    },
    "noisy_audio": {
        "path": "/path/to/Concert/noisy/chenyixun-burubujian-Concert-pro-part_0.wav",
        "array": ...,
        "sampling_rate": 44100
    }
}

Data Fields

• singer (string): The pinyin name of the singer.
• song (string): The pinyin name of the song.
• scenario (string): The acoustic scenario where the recording was captured (e.g., Concert_Hall, Arcade, Roadside, KTV).
• device (string): The type of recording device, categorized as professional (pro) or non-professional (non_pro).
• part (int): The segment number for the same song in the same scenario.
• clean_audio (Audio): The corresponding studio-quality clean vocal audio.
• noisy_audio (Audio): The noisy vocal audio recorded in the real-world scenario.

🚀 How to Use

You can easily load and use SingVERSE with the Hugging Face datasets library.

First, install the datasets library:

pip install datasets

Then, load the dataset in Python:

from datasets import load_dataset

repo_id = "amphion/SingVERSE" 
dataset = load_dataset(repo_id, split='train')

# Access a sample
sample = dataset
print(sample)

# You can directly access the audio array and sampling rate
clean_audio_array = sample['clean_audio']['array']
sampling_rate = sample['clean_audio']['sampling_rate']

📜 Citation

If you use the SingVERSE dataset in your research, please cite our paper:

@article{zhang2025singverse,
  title={{SingVERSE}: a Diverse Real-world Benchmark for Singing Voice Enhancement},
  author={Jiang, Shaohan and Zhang, Junan and Zhang, Yunjia and Yang, Jing and Fan, Fan and Wu, Zhizheng},
  journal={arXiv},
  year={2025},
  eprint={2509.20969},
  archivePrefix={arXiv}
}