sparshsinghal/deepfakeguard-finetuned

DeepfakeGuard - Fine-tuned Wav2Vec2

Model Description

This model is fine-tuned from garystafford/wav2vec2-deepfake-voice-detector for Indian voice deepfake detection.

Training Data

• 2016 training samples
• Real and AI-generated voices (Indian speakers)
• Balanced dataset (50% real, 50% fake)

Performance

• Test Accuracy: 100%
• Test F1 Score: 100%
• Confusion Matrix:
  • Real -> Real: 250
  • Fake -> Fake: 250
  • False positives: 0
  • False negatives: 0

Usage

from transformers import pipeline

pipe = pipeline('audio-classification', model='sparshsinghal/deepfakeguard-finetuned')
result = pipe('audio.wav')
print(result)

Training Details

• Base Model: garystafford/wav2vec2-deepfake-voice-detector
• Epochs: 5
• Batch Size: 8
• Learning Rate: 3e-5

License

MIT

Citation

@misc{deepfakeguard2024,
  author = {Sparsh Singhal},
  title = {DeepfakeGuard: Indian Voice Deepfake Detection},
  year = {2026},
  publisher = {Hugging Face}
}