🎵 SingMOS-Pro

[Important Notice]
We have officially released the SingMOS-Pro dataset — the official benchmark for singing voice quality assessment.

---

📚 Related Resources

• 🧾 Paper: SingMOS-Pro: A Comprehensive Benchmark for Singing Quality Assessment
  → Describes dataset design, annotation methodology, and experiments.

• 🎶 VoiceMOS 2024 Singing Track: SingMOS_v1
  → For reproducing or comparing with the official VoiceMOS 2024 track.

• 🤖 Pretrained Model: Singing MOS Predictor
  → Ready-to-use MOS prediction models trained on SingMOS and SingMOS-Pro.

---

🧩 Overview

SingMOS-Pro contains 7,981 Chinese and Japanese vocal clips, totaling 11.15 hours of singing recordings.
Most samples are recorded at 16 kHz, with a few at 24 kHz or 44.1 kHz.

This dataset enables large-scale research on singing quality assessment for tasks such as:

• Singing voice synthesis (SVS)
• Voice conversion (SVC)
• MOS prediction and correlation modeling

To use the dataset effectively, please refer to the following files:

File	Description
split.json	Defines train/test partitions
score.json	Provides system- and utterance-level MOS annotations
sys_info.json	Describes system metadata (type, model, dataset, etc.)
metadata.csv	Flat-format summary of all utterances and attributes

---

📂 Dataset Structure

SingMOS-Pro
├── wavs/                   # Singing audio clips
│   ├── sys0001-utt0001.wav
│   ├── ...
├── info/                   # Metadata and annotations
│   ├── split.json
│   ├── score.json
│   ├── sys_info.json
└── metadata.csv

---

🧾 File Descriptions

1️⃣ split.json — Dataset Partition File

Defines the train/test splits for each dataset.

Example:

{
    "dataset_name": {
        "train": ["utt0001", "utt0002", "utt0003"],
        "test": ["utt0101", "utt0102"]
    }
}

Field Descriptions:

Field	Description
dataset_name	Name of the sub-dataset (e.g., acesinger, opencpop)
train	List of utterance IDs used for training
test	List of utterance IDs used for testing

🔹 Usage: Load this file to ensure consistent dataset splits across experiments.

---

2️⃣ score.json — MOS Annotation File

Contains both system-level and utterance-level MOS (Mean Opinion Score) annotations.

Example:

{
    "system": { 
        "sys0001": {
            "score": 3.85,
            "ci": 0.07
        }
    },
    "utterance": {
        "utt0001": {
            "sys_id": "sys0001",
            "wav": "wavs/sys0001-utt0001.wav",
            "score": {
                "mos": 3.9,
                "scores": [3.5, 4.0, 4.2],
                "judges": ["J01", "J02", "J03"]
            }
        }
    }
}

Field Descriptions:

Field	Description
system	Stores system-level MOS results
sys_id	Unique system identifier (e.g., sys0001)
score	Average MOS of the system
ci	Confidence interval for the system-level MOS
utterance	Stores utterance-level annotations
utt_id	Unique utterance identifier
wav	Relative path to the audio file
score.mos	Mean MOS for the utterance
score.scores	List of individual ratings from judges
score.judges	List of judge identifiers

🔹 Usage:

• Evaluate system performance by comparing system and utterance levels.
• Compute correlations, inter-rater consistency, or build MOS prediction models.

---

3️⃣ sys_info.json — System Metadata File

Describes each singing system’s category, dataset source, model, and sampling rate.

Example:

{
    "sys0001": {
        "type": "svs",
        "dataset": "Opencpop",
        "model": "DiffSinger",
        "sample_rate": 16000,
        "tag": {
            "domain_id": "batch1",
            "other_info": "default"
        }
    }
}

Field Descriptions:

Field	Description
sys_id	Unique system identifier
type	System type: svs (singing synthesis), svc (voice conversion), or gt (ground truth)
dataset	Original dataset source
model	Model or architecture name used for generation
sample_rate	Audio sampling rate (Hz)
tag.domain_id	Batch ID or annotation domain
tag.other_info	Extra information (e.g., codec codebook, speaker transfer, etc.)

💡 "other_info": "default" means no additional metadata is available.

🔹 Usage:

• Filter systems by type or dataset.
• Analyze system-level trends and quality differences.

---

4️⃣ metadata.csv — Sample-Level Summary Table

Provides a flat-format summary of all utterances, integrating data from the JSON files. Ideal for quick indexing, filtering, and statistical analysis (e.g., via pandas).

Example:

{
    "dataset": "acesinger",
    "domain_id": 1,
    "id": "sys0001-utt0001",
    "judge_id": [1, 2, 3, 4, 5],
    "judge_lyrics_score": [],
    "judge_melody_score": [],
    "judge_score": [4.0, 4.0, 4.0, 4.0, 4.0],
    "language": "Chinese",
    "lyrics": "",
    "model_name": "ace",
    "other_info": "default",
    "raw_wav_id": "22#2100003752",
    "sample_rate": 16000,
    "split": "test",
    "system": "acesinger@ace@default",
    "system_id": "sys0001",
    "type": "svs",
    "wav": "wav/sys0001-utt0001.wav"
}

Field Descriptions:

Field	Description
dataset	Original dataset name
domain_id	Annotation batch or domain index
id	Unique utterance identifier (sysID-uttID)
judge_id	List of judge IDs who rated this utterance
judge_lyrics_score / judge_melody_score	Optional sub-dimension ratings (may be empty)
judge_score	List of overall MOS ratings from judges
language	Singing language (Chinese or Japanese)
lyrics	Transcribed lyrics text (if available)
model_name	Model or architecture name used to generate audio
other_info	Additional configuration info (e.g., codec, speaker info)
raw_wav_id	Original recording or dataset identifier
sample_rate	Sampling rate in Hz
split	Dataset partition (train / test)
system	Full system identifier (dataset@model@info)
system_id	System-level ID (matches sys_info.json)
type	System type: svs, svc, or gt
wav	Relative path to waveform file

🔹 Usage:

• Load with pandas.read_csv for analysis.
• Merge by system_id or filter by language/type.
• Perform judge-level or system-level statistical analysis.

---

🗓️ Update History

Date	Update
2025-10-09	Released SingMOS-Pro
2024-11-06	Released SingMOS
2024-06-26	Released SingMOS_v1

---

📖 Citation

If you use this dataset, please cite the following paper:

@misc{tang2025singmosprocomprehensivebenchmarksinging,
      title={SingMOS-Pro: A Comprehensive Benchmark for Singing Quality Assessment},
      author={Yuxun Tang and Lan Liu and Wenhao Feng and Yiwen Zhao and Jionghao Han and Yifeng Yu and Jiatong Shi and Qin Jin},
      year={2025},
      eprint={2510.01812},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2510.01812}
}