95,000+ audio recordings across 'fake' and 'real' speech categories are provided in four distinct preprocessing versions. The collection includes human speech from the Arctic dataset paired with synthetic versions generated by multiple text-to-speech algorithms.
Use Cases
- Train a binary classifier to distinguish between human and synthetic voices using the 'real' and 'fake' labels
- Evaluate the accuracy of detection models across the 'norm' and 'clean' dataset variants
- Identify synthetic artifacts in audio files using the 'fake' label as a ground truth
Strengths
- Contains 95,000+ audio files in .wav format
- Features four dataset variants: FoR-original, FoR-2sec, FoR-norm, and FoR-clean
- Utilizes binary labels for 'real' and 'fake' speech origins
- Includes synthetic samples generated by 6 different text-to-speech and voice conversion models