DP-VOXLET: Provable Speaker Anonymization
What it is
This paper introduces a formal definition of speaker differential privacy and a mechanism for speaker anonymization that provably satisfies this definition. It works by modifying disentangled speech representations to replace speaker properties while preserving semantic content, offering a provable lower bound on re-identification success rates for any adversary.
The problem it solves
Existing speaker anonymization systems rely on heuristic methods that lack formal privacy guarantees, leaving them vulnerable to sophisticated re-identification attacks without a provable lower bound on success rates.
What you could use it for
- Anonymizing speech data for privacy-preserving machine learning training
- Creating synthetic speech datasets with guaranteed privacy bounds
- Compliance with privacy regulations requiring provable anonymization guarantees
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.