LibriVoc Dataset

Overview

LibriVoc is an open-source, large-scale dataset designed for vocoder artifact detection in synthesized speech. Built upon the LibriTTS speech corpus—a widely-used resource in text-to-speech research derived from LibriVox audiobooks—LibriVoc provides a comprehensive benchmark for evaluating and detecting synthetic speech artifacts.

Dataset Access: Download from Google Drive

Dataset Composition

LibriVoc features speech samples generated using six state-of-the-art neural vocoders across three major categories:

• Autoregressive vocoders: WaveNet, WaveRNN
• GAN-based vocoders: MelGAN, Parallel WaveGAN
• Diffusion-based vocoders: WaveGrad, DiffWave

Dataset Statistics

• Real samples (training): 126.41 hours
• Synthesized samples (training): 118.08 hours
• Total samples: 43,809 (Training: 33,236 | Development: 5,736 | Testing: 4,837)

Audio Distribution by Vocoder

Model	train-clean-100	train-clean-360	dev-clean	test-clean
WaveNet	4.28	15.49	0.75	0.76
WaveRNN	4.33	14.92	0.67	0.72
MelGAN	4.36	15.26	0.71	0.76
Parallel WaveGAN	4.37	15.54	0.68	0.75
WaveGrad	4.19	15.81	0.76	0.74
DiffWave	4.16	15.37	0.62	0.66
Total	25.69	92.39	4.19	4.39

All values represent hours of audio

Methodology

Self-Vocoding Process

Each vocoder synthesizes waveform samples from mel spectrograms extracted from original LibriTTS samples—a process called "self-vocoding." By providing identical mel spectrograms to all vocoders, any unique artifacts in the synthesized samples can be directly attributed to the specific vocoder used for reconstruction.

Dataset Design

To prevent overfitting to speaker identity, the dataset employs a balanced speaker-level distribution:

1. 25% of speakers: Only real (original) samples
2. 25% of speakers: Only synthesized samples
3. 50% of speakers: Mixed composition (50% real, 50% synthesized per speaker)

This stratified design ensures that classifiers trained on LibriVoc learn to detect vocoder artifacts rather than memorizing speaker characteristics.

Data Splits

The dataset is divided into three non-overlapping sets:

• Training set: 33,236 samples
• Development set: 5,736 samples
• Testing set: 4,837 samples

Applications

LibriVoc is suitable for:

• Training and evaluating vocoder artifact detectors
• Benchmarking synthetic speech detection algorithms
• Researching vocoder-specific acoustic signatures
• Developing robust deepfake audio detection systems

Citation

If you use the LibriVoc dataset in your research, please cite the original paper: https://link.springer.com/chapter/10.1007/978-3-031-49803-9_11

@incollection{sun2023using,
  title={Using Vocoder Artifacts For Audio Deepfakes Detection},
  author={Sun, Chengzhe and AlBadawy, Ehab and Davison, Timothy F and Robinson, Sarah R and Chang, Ming-Ching and Lyu, Siwei},
  booktitle={Adversarial Multimedia Forensics},
  pages={263--282},
  year={2023},
  publisher={Springer}
}

License

Print ISBN 978-3-031-49802-2

Online ISBN 978-3-031-49803-9