NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·1h ago
AI Research
arXiv CS.AI·Jun 10·5 min read
Dual-Branch Gated Fusion for Open-Set Audio Deepfake Source Tracing
TL;DR
Attributing synthetic audio to its source is challenging because existing models struggle with unseen synthesizers. A dual-branch gated fusion framework was developed, combining XLSR-53, a speech representation model, with CORES, a comprehensive audio descriptor.
✦ Why It Matters
Engineers can leverage this framework to improve audio deepfake detection systems, enhancing their robustness against unseen synthesizers.
Key Takeaways
- Utilize the dual-branch gated fusion framework for audio deepfake tracing.
- Combine XLSR-53 with CORES for improved artifact detection.
- Leverage input-conditioned gating for balanced model training.
- Achieve high accuracy and reduced false positives in open-set scenarios.
How It Works
The dual-branch gated fusion framework operates by integrating two models: XLSR-53, which excels in recognizing known audio patterns, and CORES, which captures a broader range of synthesis artifacts. The input-conditioned gate dynamically adjusts the influence of each model based on the input data, allowing for more accurate predictions in diverse scenarios.
This adaptive weighting is crucial for effectively distinguishing between in-domain and out-of-domain audio sources.