ArshTandon/AAT_Deekfake_and_real_Audio

Audio Deepfake Detection Dataset

Overview

This dataset provides one of the most comprehensive audio deepfake detection resources available, consisting of real and synthetic audio samples standardized for machine learning research. It is designed to support development and benchmarking of deepfake detection algorithms.

Dataset Composition

• Total audio files: 189,221

  • Real: 101,172 (53.5%)
  • Fake: 88,049 (46.5%)
  • Class ratio (Real:Fake): 1.15:1

• Total duration: 315.58 hours

  • Real: 169.24 hours (53.6%)
  • Fake: 146.35 hours (46.4%)
  • Duration ratio: 1.16:1

• Average segment length: 6.00 seconds

  • Real mean: 6.02 ± 3.12 s
  • Fake mean: 5.98 ± 3.67 s

Data Sources and Selection Criteria

We curated our dataset from four primary sources to ensure diversity across language, modality, acoustic environments, and synthesis methods:

MLAAD (Multi-Language Audio Anti-Spoofing Dataset): A multilingual benchmark spanning 35+ languages and 59–91 TTS systems across concatenative, parametric, and neural architectures. Read speech segments chunked to 5–10 s, standardized to 22.05 kHz and 16-bit PCM. Manual auditing and SNR analysis ensured quality.

FakeAVCeleb (Audio-only Subset): Deepfake audio extracted from video deepfakes with cross-modal lip-sync artifacts. Contains ∼5,000 fake and 1,000 real 5–10 s segments (50% overlap), preserving compression artifacts.

YouTube Long-Form Real Speech: 31,403 authentic 5–10 s segments from speeches, interviews, and podcasts. Captures natural prosody, disfluencies, and acoustic variation across professional and mobile recordings.

YouTube Gaming Series Fake Commentary: 19,596 synthetic commentary segments in noisy gaming environments. Includes voice cloning of influencers, background audio, and dynamic expressions. Segments chunked to 5–10 s with VAD and noise profiling.

These sources were selected to cover linguistic diversity, TTS technology variety, recording conditions, modality, and real-world application scenarios, then balanced to form the final dataset.

Audio Quality Metrics (Sample of 500 files per class)

Metric	Real Audio	Fake Audio
Mean duration	6.02 s	5.98 s
Median duration	5.77 s	4.95 s
Duration range	0.70–15.93 s	0.82–23.32 s
Standard deviation	3.12 s	3.67 s
95th percentile	11.85 s	13.43 s

Format Standardization

All files have been converted and verified to:

• Sample rate: 16 kHz
• Channels: Mono (1 channel)
• Bit depth: 16-bit PCM (linear)
• Container: WAV (RF64 compatible)

Original Format Distribution

• Sample rates: 8 kHz (3.7%), 16 kHz (61.2%), 22.05 kHz (14.8%), 44.1 kHz (19.1%), 48 kHz (1.2%)
• Channels: Mono (82.3%), Stereo (17.7%)
• Bit depths: 16-bit PCM (76.4%), 24-bit PCM (5.2%), 32-bit float (9.1%), MP3 (7.8%), Other (AAC/OGG, 1.5%)

Feature Extraction

For each file, 34 acoustic features were extracted using Librosa and PyTorch Audio:

• Spectral (26 features): 13 MFCC coefficients + their means and standard deviations; spectral centroid, bandwidth, rolloff, and contrast.
• Temporal/Energy (4 features): Zero-crossing rate, RMS energy (mean, std, dynamic range).
• Voice Quality (4 features): Fundamental frequency mean and std (YIN algorithm); harmonic-to-noise ratio; spectral flatness.

Preprocessing steps included VAD-based silence removal, DC offset filtering, pre-emphasis (0.97 coefficient), and normalization.

Class Balancing Strategy

• Maintained slight real-over-fake imbalance (53.5% vs. 46.5%) to mirror real-world distributions.
• Employed class weighting, stratified sampling, and balanced batch generation during model training.

Source-Type Distribution

Source Category	Real	Fake
Read Speech	35,410 (35.0%)	33,217 (37.7%)
Conversational	41,480 (41.0%)	36,741 (41.7%)
Entertainment	24,282 (24.0%)	18,091 (20.6%)

Recording Quality Distribution

Quality Level	Real	Fake
Studio	42,492 (46.7%)	45,786 (46.7%)
Semi-professional	31,363 (31.5%)	28,176 (31.5%)
Amateur/Field	27,317 (21.8%)	14,087 (21.8%)

Usage Instructions

1. Partitioning: Use 80% of files for training, 10% for validation, and 10% for testing, stratified by class and source type.
2. Loading: Audio files are organized in real/ and fake/ directories. Metadata files provide labels and sampling info.
3. Feature Vectors: Precomputed NPZ files of feature matrices are available under /features/ for quick model prototyping.

Citation

Citation

If you use this dataset, please cite:

Tandon, A. (2025). AAT_Deekfake_and_real_Audio [Dataset]. Hugging Face. Available at: https://huggingface.co/datasets/ArshTandon/AAT_Deekfake_and_real_Audio

License

This dataset is released under the Apache License 2.0. See LICENSE for details.