Continual Speaker Identity Unlearning with Minimal Interference (arxiv.org)

Researchers have introduced a groundbreaking framework called Cumulative ORThogonal Identity Suppression (CORTIS) that addresses a significant gap in the field of machine unlearning, specifically within zero-shot text-to-speech (ZS-TTS) systems. While previous methods presumed that all speaker identity unlearning requests would occur simultaneously, CORTIS recognizes that removal requests often come sequentially. This innovation is crucial as existing state-of-the-art techniques inadvertently revive previously unlearned speakers when new unlearning requests are implemented, thereby compromising user privacy—a core goal of machine unlearning. CORTIS employs a sophisticated combination of Fisher-information-based parameter masking and orthogonal projection techniques, allowing it to efficiently target updates to speaker-specific weights without requiring access to data from previously unlearned speakers. This approach ensures that every new unlearning request maintains the erasure of prior speaker identities over extended sequences of requests. By significantly outperforming traditional methods, CORTIS marks a significant advancement for the AI/ML community, particularly in enhancing privacy protocols in voice synthesis technology. A public demo of this innovative system is currently available for demonstration.