NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·2h ago
AI Research
arXiv CS.AI·Jun 11·4 min read
Feature-Aligned Speech Watermarking for Robustness to Reconstruction Distortions
TL;DR
Researchers identified a gap in LLM evaluation benchmarks. They built a synthetic dataset with 10k adversarial prompts targeting reasoning failures.
✦ Why It Matters
Use this benchmark to audit LLM robustness before deploying in production reasoning pipelines.
Key Takeaways
- Explore feature-aligned watermarking for better audio security.
- Utilize pretrained speech codecs for watermark generation.
- Optimize watermark embedding using VAD and perceptual losses.
- Test robustness against various speech reconstruction models.
How It Works
The feature-aligned watermarking method aligns the watermark with the original speech features, allowing for higher energy levels. This is achieved by generating a pseudo-speech watermark using a pretrained speech codec and embedding it into the audio's spectrogram.
The use of VAD loss ensures that the watermark is placed in regions where speech is present, while perceptual losses help maintain audio quality.