SatyaDrishti Audio Deepfake Detector
ΰ€Έΰ€€ΰ₯ΰ€―ΰ€¦ΰ₯ΰ€·ΰ₯ΰ€ΰ€Ώ β AI-Powered Digital Forensics Platform
Chandigarh Police National Hackathon 2026 β Track 4
Architecture
- Backbone: facebook/wav2vec2-base (self-supervised speech)
- Sequence: BiGRU (256 hidden x 2 layers, bidirectional)
- Attention: Multi-Head Attention (4 heads)
- Head: LayerNorm -> Dropout -> Linear(512)->GELU -> Linear(256)->GELU -> Linear(2)
- Input: 16kHz mono audio, 4 seconds
Training Datasets (~50K+ samples)
- CommonVoice 13.0 β 15K real human speech
- LibriSpeech β 5K real read speech
- MLAAD β 10K synthetic (175 TTS models)
- FakeOrReal β 8K balanced
- Voice Conversion samples β 5K
- Kokoro TTS β 5K generated
Performance
- Val AUC: 1.0000
- Classes: REAL (0), FAKE/SYNTHETIC (1)
Companion Model
satyadrishti-deepfake-detector β image/video