Hugging Face's logo

Datasets:
XuepingZhang
/
ESDD2-CompSpoof-V2

Modalities:
Audio
Text
Formats:
webdataset
Size:
10K - 100K
Libraries:
Datasets
WebDataset
License:
Dataset card Data Studio Files
xet
Community
1

Acknowledge license to accept the repository

This dataset is a derived dataset constructed by combining and mixing audio samples from multiple publicly available datasets. - The MLAAD and VCapAV datasets are released under the CC BY-NC 4.0 license. - The LibriTTS, EnvSDD and VGGSound datasets is released under the CC BY 4.0 license. - The Common Voice dataset is released under the Creative Commons CC0 1.0 Universal license. - The ASVspoof 5 dataset is released under the ODC-By License. - The english-conversation-corpus dataset is released under the GPLv3License. - The AudioCaps dataset is released under the mit License. - The TUTASC, TUTSED and UrbanSound datasets are released under the Non-Commercial License.

Users must comply with the license terms of each original dataset. The authors do not claim ownership of the original audio content. Due to the inclusion of datasets licensed under CC BY-NC 4.0 license, this dataset is released under the CC BY-NC 4.0 license.

Log in or Sign Up to review the conditions and access this dataset content.

πŸŽ™οΈ CompSpoof V2 Dataset

1. Introduction

CompSpoof V2 is a dataset designed for component-level anti-spoofing detection research, where either the speech or the environmental sound component (or both) may be spoofed.

πŸ“Š CompSpoof V2 contains over 250k audio samples, with a total duration of approximately 283 hours. ⏱️ Each audio sample has a fixed length of 4 seconds and is provided at multiple sampling rates, enabling a more faithful simulation of real-world acoustic and system-level variations.

Building upon CompSpoof dataset, CompSpoof V2 significantly expands the diversity of attack sources, environmental sounds, and mixing strategies. ✨ In addition, newly generated audio samples are distributed across the test set and are specifically designed to serve as detection data under unseen conditions. πŸ”Š To further simulate realistic transmission effects, portions of the test set are processed by audio codec transformation.

πŸ€— CompSpoof V2 Download Link: https://huggingface.co/datasets/XuepingZhang/ESDD2-CompSpoof-V2/

πŸ’» Baseline code: https://github.com/XuepingZhang/ESDD2-Baseline

2. Audio Class Description and Samples

Below are audio samples from the CompSpoof V2 dataset. For each class, we provide the mixed/original audio, along with the speech and environment sources.

Class 0 β€” Original

Label: original

Description: Original bona fide speech and corresponding environment audio without mixing

Class 1 β€” Bona fide + Bona fide

Label: bonafide_bonafide

Description: Bona fide speech mixed with another bona fide environmental audio

Class 2 β€” Spoofed Speech + Bona fide Environment

Label: spoof_bonafide

Description: Spoof speech mixed with bona fide environmental audio

Class 3 β€” Bona fide Speech + Spoofed Environment

Label: bonafide_spoof

Description: Bona fide speech mixed with spoof environmental audio

Class 4 β€” Spoofed Speech + Spoofed Environment

Label: spoof_spoof

Description: Spoof speech mixed with spoof environmental audio

3. CompSpoof V2 VS CompSpoof

CompSpoof dataset is our previously released dataset designed for component-level spoofing detection. Building upon this foundation, we introduce CompSpoof V2, a substantially upgraded version with expanded task formulation. The key differences between CompSpoof and CompSpoof V2 are summarized below.

Aspect CompSpoof CompSpoof V2
Data volume 2.5k audio clips, about 7 hours πŸ“Š more than 250k audio clips, about 283 hours
Data sources SSTC, ASV5, VggSound, VcapAV, Common Voice AudioCaps, VggSound, CommonVoice, LibriTTS, english-conversation-corpus,ASV5, MLAAD,TUTASC, TUTSED, UrbanSound, VGGSound, EnvSDD, VcapAV
Duration range from 5 to 21 seconds ⏱️ 4 seconds/audio clip
Codec transformation ❌ βœ…
Newly generated audio ❌ βœ…

4. Dataset Structure

The dataset follows a hierarchical directory structure organized by data split:

CompSpoof
β”œβ”€β”€ development                     # training and val data，including audio source
β”‚   β”œβ”€β”€ env_source                  # environmental sound audio used as the environmental sound component in the mixture
β”‚   β”œβ”€β”€ metadata                    # metadata of development set
β”‚       β”œβ”€β”€ train.csv
β”‚       └── val.csv
β”‚   β”œβ”€β”€ mixed_audio                 # mixed audio files, which **don't** belong to the `original` class
β”‚   β”œβ”€β”€ original_audio              # audios belong to `original` class
β”‚   └── speech_sources              # speech audio used as the speech component in the mixture
β”‚
β”œβ”€β”€ eval                            # eval set data, without audio source
β”‚   β”œβ”€β”€ audio                       # audio files
β”‚   └── metadata                    # metadata of eval set, which only has file name
β”‚       └── eval.csv
β”‚
β”œβ”€β”€ eval_source (Released later)    # eval set data，including audio source
β”‚   β”œβ”€β”€ env_sources                 # environmental sound audio used as the environmental sound component in the mixture
β”‚   β”œβ”€β”€ metadata                    # metadata of eval set, with full annotation
β”‚   β”‚   └── eval.csv
β”‚   β”œβ”€β”€ mixed_audio                 # mixed audio files, which **don't** belong to the `original` class
β”‚   β”œβ”€β”€ original_audio              # audios belong to `original` class
β”‚   └── speech_sources              # speech audio used as the speech component in the mixture
|
β”œβ”€β”€ test                            # test set data, without audio source
β”‚   β”œβ”€β”€ audio                       # audio files
β”‚   └── metadata                    # metadata of test set, which only has file name
β”‚       └── test.csv
β”‚
└── test_source (Released later)    # test set audio data，including audio source
    β”œβ”€β”€ env_sources                 # environmental sound audio used as the environmental sound component in the mixture
    β”œβ”€β”€ metadata                    # metadata of test set, with full annotation
    β”‚   └── test.csv
    β”œβ”€β”€ mixed_audio                 # mixed audio files, which **don't** belong to the `original` class
    β”œβ”€β”€ original_audio              # audios belong to `original` class
    └── speech_sources              # speech audio used as the speech component in the mixture

5. Audio Source

The audio sources for each category are as follows:

πŸ‹οΈ train & val set

Label Original Source Speech Source Environmental Sound Source
original AudioCaps, VggSound - -
bonafide_bonafide - CommonVoice, LibriTTS, english-conversation-corpus AudioCaps, TUTASC, TUTSED, UrbanSound, VGGSound
bonafide_spoof - CommonVoice, LibriTTS EnvSDD, VcapAV
spoof_bonafide - ASV5, MLAAD AudioCaps, TUTASC, TUTSED, UrbanSound, VGGSound
spoof_spoof - ASV5, MLAAD EnvSDD, VcapAV

🏁 eval & test set

Label original source speech source environmental sound source
original AudioCaps, VggSound - -
bonafide_bonafide - CommonVoice, LibriTTS, english-conversation-corpus AudioCaps, TUTASC, TUTSED, UrbanSound, VGGSound
bonafide_spoof - CommonVoice, LibriTTS EnvSDD, VcapAV, New Generated
spoof_bonafide - ASV5, MLAAD, New Generated AudioCaps, TUTASC, TUTSED, UrbanSound, VGGSound
spoof_spoof - ASV5, MLAAD, New Generated EnvSDD, VcapAV, New Generated

6. Data Splits

The dataset is divided into three standard splits:

  • Training set: used for model training
  • Validation set: used for validation and hyper-parameter tuning
  • Eval & Test set: used for final performance reporting

Training set and validation set have the same date source and class distribution.

Eval set and Test set share the same date source and class distribution. Eval set and Test set share some new generated audios which are unseen in training and validation set. Portions of the Eval set and Test set have been processed with audio codec toolkits.

The quantity and proportion of audios for each category in each set are as follows:

πŸ“Š train set (Total: 175361)

Label Count Ratio
bonafide_spoof 50361 28.72%
original 48639 27.74%
spoof_spoof 29413 16.77%
bonafide_bonafide 25189 14.36%
spoof_bonafide 21759 12.41%

πŸ“Š val set (Total: 24864)

Label Count Ratio
bonafide_spoof 8071 32.46%
original 6939 27.91%
spoof_spoof 4657 18.73%
bonafide_bonafide 2784 11.20%
spoof_bonafide 2413 9.70%

πŸ“Š eval set (Total: 27605)

Label Count Ratio
bonafide_spoof 7655 27.73%
original 7455 27.01%
spoof_spoof 5945 21.54%
bonafide_bonafide 3570 12.93%
spoof_bonafide 2980 10.80%

πŸ“Š test set (Total: 27603)

Label Count Ratio
bonafide_spoof 7672 27.79%
original 7415 26.86%
spoof_spoof 5894 21.35%
bonafide_bonafide 3635 13.17%
spoof_bonafide 2987 10.82%

7. Metadata

πŸ—‚οΈ Metadata is provided in CSV format, with one row per audio file. Each field describes the source, generation process, and mixing configuration of the corresponding composite spoofing sample.

The meaning of each field in Metadata is as follows:

  • audio_path: Relative path to the final mixed audio file used for training or evaluation.

  • label: Class label of the audio sample. Typical values include: original, bonafide_bonafide, spoof_bonafide, bonafide_spoof, spoof,spoof

  • split: Dataset split indicator: train, val, eval, test

  • original_audio_source: Source dataset of the original audio, e.g., AudioCaps.

  • speech_path: Path to the speech signal used as the speech component in the mixture.

  • speech_source: Source dataset of the speech signal, e.g., ASV5, CommonVoice.

  • speech_generation_mothed: Generation method used to produce the speech signal, e.g., TTS (text-to-speech), VC (voice-conversion).

  • speech_generation_source: Dataset to generate the spoofed speech, e.g., a spoofed speech is generated by TTS, the text for generation is the source.

  • speech_generation_model: Model used to generate the spoofed speech.

  • env_path: Path to the environmental sound used as the environmental sound component in the mixture.

  • env_source: Source dataset of the environmental sound, e.g., EnvSDD, VcapAV.

  • env_generation_mothed: Method used to generate the spoofed environmental sound, e.g.,TTA (text-to-audio).

  • env_generation_source: Dataset to generate the spoofed environmental sound, e.g., a spoofed speech is generated by TTS, the text for generation is the source.

  • env_generation_model:Model used to generate the spoofed environmental sound.

  • mix_target_snr: Target signal-to-noise ratio (SNR, in dB) used when mixing the speech and environmental sound.

8. Citation

πŸ“š If you use CompSpoof V2 in your research, please cite the corresponding paper:

@dataset{zhang2025esdd2compspoofv2,
  title     = {ESDD2-CompSpoof-V2: A Composite Spoofing Dataset for Speech Anti-Spoofing},
  author    = {Zhang, Xueping and Li, Ming},
  year      = {2025},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/XuepingZhang/ESDD2-CompSpoof-V2}
}

9. License πŸ”

This dataset is a derived dataset constructed by combining and mixing audio samples from multiple publicly available datasets.

  • The MLAAD and VCapAV datasets are released under the CC BY-NC 4.0 license.
  • The LibriTTS, EnvSDD and VGGSound datasets is released under the CC BY 4.0 license.
  • The Common Voice dataset is released under the Creative Commons CC0 1.0 Universal license.
  • The ASVspoof 5 dataset is released under the ODC-By License.
  • The english-conversation-corpus dataset is released under the GPLv3License.
  • The AudioCaps dataset is released under the mit License.
  • The TUTASC, TUTSED and UrbanSound datasets are released under the Non-Commercial License.

Users must comply with the license terms of each original dataset. The authors do not claim ownership of the original audio content. Due to the inclusion of datasets licensed under CC BY-NC 4.0 license, this dataset is released under the CC BY-NC 4.0 license.

10. Contact Information

For questions, issues, or collaboration inquiries, please contact:

Downloads last month
22