Audio Avatar Fingerprinting: An Approach for Authorized Use of Voice Cloning in the Era of Synthetic Audio
New method detects if a synthetic voice was generated by its legitimate owner, not an imposter.
A new research paper titled 'Audio Avatar Fingerprinting: An Approach for Authorized Use of Voice Cloning in the Era of Synthetic Audio' tackles a critical emerging problem in AI speech synthesis. Authored by Candice R. Gerstner, the work introduces the concept of 'audio avatar fingerprinting'—a forensic task designed not just to detect fake speech, but to verify whether a synthetic voice clip was generated by an authorized user. This is crucial as AI voice cloning technology, which can create realistic audio from just seconds of reference speech, creates new risks for authentication systems and broadcasting while simultaneously enabling beneficial applications like audio enhancement for communication.
The research represents a first-of-its-kind experimentation by adapting an existing, general-purpose speaker verification model for the dual tasks of fake speech detection and authorized use verification. A key contribution is the creation of a novel dataset specifically for this task, filling a significant gap as no prior dataset allowed for testing the verification of authorized synthetic audio. This work lays essential groundwork for future security systems that must distinguish between malicious voice deepfakes and legitimate, user-driven synthetic speech, enabling safer adoption of voice AI tools.
- Introduces 'audio avatar fingerprinting,' a new forensic task to verify if an AI-generated voice is authorized by its owner.
- Adapts an off-the-shelf speaker verification model for fake speech detection and authorization verification, a novel application.
- Creates and releases a new dataset specifically for testing authorized synthetic audio verification, addressing a prior data gap.
Why It Matters
Enables secure, legitimate use of voice cloning for business and communication while providing a tool to combat malicious deepfakes.