README.md

metadata

language:
  - en
license: apache-2.0
size_categories:
  - 10K<n<100K
task_categories:
  - audio-classification
pretty_name: 'SINE: Speech INfilling Edit Dataset'
tags:
  - audio
  - speech
  - deepfake-detection
configs:
  - config_name: preview
    data_files:
      - split: train
        path: preview/train-*
dataset_info:
  config_name: preview
  features:
    - name: audio
      dtype:
        audio:
          sampling_rate: 16000
    - name: filename
      dtype: string
    - name: category
      dtype: string
    - name: timestamp
      dtype: string
    - name: label
      dtype: int64
    - name: manipulation_type
      dtype: string
  splits:
    - name: train
      num_bytes: 10309938
      num_examples: 30
  download_size: 10039423
  dataset_size: 10309938

SINE Dataset

Overview

The Speech INfilling Edit (SINE) dataset is a comprehensive collection for speech deepfake detection and audio authenticity verification. This dataset contains ~87GB of audio data distributed across 32 splits, featuring both authentic and synthetically manipulated speech samples.

Dataset Statistics

• Total Size: ~87GB
• Number of Splits: 32 (split-0.tar.gz to split-31.tar.gz)
• Audio Format: WAV files
• Source: Speech edited from LibriLight dataset with transcripts obtained from LibriHeavy

Audio Statistics

Audio Types	Subsets	# of Samples	# of Speakers	Durations (h)	Audio Lengths (s)
					min	max
Real/Resyn	train	26,547	70	51.82	6.00	8.00
	val	8,676	100	16.98	6.00	8.00
	test	8,494	900	16.60	6.00	8.00
Infill/CaP	train	26,546	70	51.98	5.40	9.08
	val	8,686	100	16.99	5.45	8.76
	test	8,493	903	16.64	5.49	8.85

Data Structure

Each split (e.g., split-0/) contains:

split-X/
├── combine/                    # Directory containing all audio files (~11,076 files)
│   ├── dev_real_medium-*.wav          # Authentic audio samples
│   ├── dev_edit_medium-*.wav          # Edited audio samples
│   ├── dev_cut_paste_medium-*.wav     # Cut-and-paste manipulated samples
│   └── dev_resyn_medium-*.wav         # Resynthesized audio samples
├── medium_real.txt             # Labels for authentic audio (2,769 entries)
├── medium_edit.txt             # Labels for edited audio (2,769 entries)
├── medium_cut_paste.txt        # Labels for cut-paste audio (2,769 entries)
└── medium_resyn.txt            # Labels for resynthesized audio (2,769 entries)

Audio Categories

1. Authentic Speech (dev_real_medium-*)

• Original, unmodified speech recordings from LibriVox audiobooks
• Labeled as class 1 (authentic)
• Simple time annotation format: filename start-end-T label

2. Resynthesized Speech (dev_resyn_medium-*)

• Speech regenerated from mel-spectrogram using HiFi-GAN vocoder
• Labeled as class 1 (authentic)
• Simple time annotation format

3. Edited Speech (dev_edit_medium-*)

• Audio samples with artificial modifications/edits
• Labeled as class 0 (manipulated)
• Complex time annotation with T/F segments indicating real/fake portions

4. Cut-and-Paste Speech (dev_cut_paste_medium-*)

• Audio created by cutting and pasting segments from different sources
• Labeled as class 0 (manipulated)
• Complex time annotation showing spliced segments

Label Format

Simple Format (Real/Resyn)

filename start_time-end_time-T label

Example:

dev_real_medium-100-emerald_city_librivox_64kb_mp3-emeraldcity_02_baum_64kb_21 0.00-7.92-T 1

Complex Format (Edit/Cut-Paste)

filename time_segment1-T/time_segment2-F/time_segment3-T label

Example:

dev_edit_medium-100-emerald_city_librivox_64kb_mp3-emeraldcity_02_baum_64kb_21 0.00-4.89-T/4.89-5.19-F/5.19-8.01-T 0

Where:

• T = True/Authentic segment
• F = False/Manipulated segment
• label: 1 = Authentic, 0 = Manipulated

Applications

This dataset is suitable for:

• Speech Deepfake Detection: Binary classification of authentic vs. manipulated speech
• Temporal Localization: Identifying specific time segments that contain manipulations
• Manipulation Type Classification: Distinguishing between different types of audio manipulation
• Robustness Testing: Evaluating detection systems across various manipulation techniques

Citation

This is a joint work done by NVIDIA and National Taiwan University. If you use this dataset, please cite:

@inproceedings{huang2024detecting,
  title={Detecting the Undetectable: Assessing the Efficacy of Current Spoof Detection Methods Against Seamless Speech Edits},
  author={Huang, Sung-Feng and Kuo, Heng-Cheng and Chen, Zhehuai and Yang, Xuesong and Yang, Chao-Han Huck and Tsao, Yu and Wang, Yu-Chiang Frank and Lee, Hung-yi and Fu, Szu-Wei},
  booktitle={2024 IEEE Spoken Language Technology Workshop (SLT)},
  pages={652--659},
  year={2024},
  organization={IEEE}
}

License

This dataset is released under the Apache 2.0 License.

---

Note: This dataset is intended for research purposes in speech authenticity verification and deepfake detection. Please use responsibly and in accordance with applicable laws and regulations.