Environmental sound deepfake detection using EAT and AASIST
Python
2
9 commits
updated Feb 13, 2026
Generative AI can now create highly realistic environmental sounds, such as sirens, footsteps, and nature amibence, raising concerns about misinformation and safety. Unlike speech deepfakes, environmental sounds span diverse categories and exhibit complex generation artifacts.
This repo contains a robust Deepfake Detection System designed to distinguish between real recordings and AI-generated audio. Our solution leverages Self-Supervised Learning (SSL) to generalize effectively to unseen audio generators (Text-to-Audio and Audio-to-Audio models).
We utilize the EnvSDD dataset, a large-scale curated collection designed for benchmarking environmental deepfake detection.
Our architecture combines a powerful pre-trained front-end with a graph-based back-end to capture subtle deepfake artifacts:
The model was evaluated using the Equal Error Rate (EER) metric, where a lower score indicates better performance.
| Metric | Score |
|---|---|
| Baseline AASIST EER | ~15.02% |
| Baseline W2V2+AASIST EER | ~48.00% |
| Baseline BEATs+AASIST EER | ~13.20% |
| Proposed EAT+AASIST EER | ~2.48% |
By combining EAT with AASIST, our solution achieves state-of-the-art performance, drastically reducing error rates compared to standard baselines.
Environmental sound deepfake detection using EAT and AASIST
Python
2
9 commits
updated Feb 13, 2026
Generative AI can now create highly realistic environmental sounds, such as sirens, footsteps, and nature amibence, raising concerns about misinformation and safety. Unlike speech deepfakes, environmental sounds span diverse categories and exhibit complex generation artifacts.
This repo contains a robust Deepfake Detection System designed to distinguish between real recordings and AI-generated audio. Our solution leverages Self-Supervised Learning (SSL) to generalize effectively to unseen audio generators (Text-to-Audio and Audio-to-Audio models).
We utilize the EnvSDD dataset, a large-scale curated collection designed for benchmarking environmental deepfake detection.
Our architecture combines a powerful pre-trained front-end with a graph-based back-end to capture subtle deepfake artifacts:
The model was evaluated using the Equal Error Rate (EER) metric, where a lower score indicates better performance.
| Metric | Score |
|---|---|
| Baseline AASIST EER | ~15.02% |
| Baseline W2V2+AASIST EER | ~48.00% |
| Baseline BEATs+AASIST EER | ~13.20% |
| Proposed EAT+AASIST EER | ~2.48% |
By combining EAT with AASIST, our solution achieves state-of-the-art performance, drastically reducing error rates compared to standard baselines.
