Chanson-0803/SinTechSVS-singing-technique-annotation

SinTechSVS: Singing Technique Annotation

A phoneme-level manual annotation dataset for singing techniques, covering two dimensions: pitch technique and timbral technique. This annotation was produced as part of the SinTechSVS project (NUS Sound and Music Computing Lab), based on the paper SinTechSVS: A Singing Technique Controllable Singing Voice Synthesis System, published in IEEE/ACM TASLP 2024.

• 📄 Paper: PDF
• 🎧 Demo: https://danny-nus.github.io/SinTechSVS/

Dataset Overview

• Number of samples: 500
• Annotation dimensions: each sample contains two (in principle equal-length) label sequences
  • pitch_technique_seq: pitch-dimension technique sequence
  • timbral_technique_seq: timbre-dimension technique sequence
• Associated audio: this dataset contains annotations only, no audio files. The id field is a 10-digit numeric identifier (e.g. 2001000002) corresponding to the same-named wav file in the segments/wavs/ directory of the Opencpop dataset. The original paper explicitly states that these annotations were manually produced on the Opencpop dataset (by three annotators with professional music backgrounds). Audio must therefore be obtained separately from the official Opencpop release, subject to its TERMS_OF_ACCESS.

Field Description

Field	Type	Description
id	int64	Sample ID, corresponding to the original audio segment ID
source_dataset	string	Source dataset of the underlying audio ("Opencpop", see above)
pitch_technique_seq	list[string]	Pitch technique label sequence
timbral_technique_seq	list[string]	Timbral technique label sequence
seq_length	int64	Length (number of tokens) of pitch_technique_seq
pitch_timbre_length_match	bool	Whether the two sequences have matching lengths (see "Known Issues")

Label Definitions

Each token in a sequence is either a numeric string (label ID) or a special symbol. The mapping from numeric ID to technique name is also provided in label_maps.json:

Pitch technique

ID	Technique
0	scooping
1	bend
2	drop
3	hiccup
4	melisma
5	straight

Timbral technique

ID	Technique
0	vocal fry
1	falsetto
2	breathy
3	whisper
4	belting
5	regular

Special symbols

Symbol	Meaning
SP	silence / pause
AP	aspiration / breath

Label digits typically appear in pairs (likely corresponding to the start/end frame of a labeled segment, or a repeated confirmation). Please refer to the original annotation pipeline to confirm the exact granularity (frame-level vs. phoneme-level).

Known Issues

• Sample id=2021000819 has a length mismatch between pitch_technique_seq (16 tokens) and timbral_technique_seq (15 tokens), likely due to a missing frame in the timbral annotation. This is flagged via pitch_timbre_length_match=False and has not been corrected or removed — please handle it according to your use case.
• One occurrence of an extraneous whitespace in the original spreadsheet (which caused incorrect tokenization) has been fixed. This is unrelated to the issue above.
• When training the singing technique classifier, the original paper excluded the whisper (timbre) and hiccup (pitch) classes due to insufficient label counts, retaining only 8 classes for model training and evaluation. This dataset (the 500-sample manually annotated subset) still fully retains the original annotations for these two labels; no classes have been removed.

Loading

from datasets import load_dataset

ds = load_dataset("<your-hf-username>/SinTechSVS-singing-technique-annotation")
print(ds["train"][0])

File Structure

.
├── README.md
├── data/
│   └── train-00000-of-00001.parquet   # Main data (standard HF datasets format)
├── annotations.jsonl                   # Same data in JSONL format, for easy inspection/grep
└── label_maps.json                     # Mapping from pitch/timbre label IDs to names

Citation

@article{zhao2024sintechsvs,
  title={Sintechsvs: A singing technique controllable singing voice synthesis system},
  author={Zhao, Junchuan and Chetwin, Low Qi Hong and Wang, Ye},
  journal={IEEE/ACM Transactions on Audio, Speech, and Language Processing},
  volume={32},
  pages={2641--2653},
  year={2024},
  publisher={IEEE}
}

Use of this dataset should also comply with the terms of the Opencpop dataset (please obtain audio through the official Opencpop release).

License

To be determined. Please confirm the release license for this annotation data via the paper, project page, or by contacting the authors (junchuan@comp.nus.edu.sg), and refer to the TERMS_OF_ACCESS of the Opencpop dataset.