Continual Speaker Identity Unlearning with Minimal Interference

Researchers introduced CORTIS, a continual speaker identity unlearning method for zero-shot text-to-speech.

The paper finds that existing unlearning methods assume all deletions happen at once, which can let later requests restore speakers removed earlier.

CORTIS combines Fisher-information-based parameter masking with orthogonal projection to keep prior removals intact during new unlearning requests.

The approach targets a practical privacy gap in models like VoiceBox by supporting sequential, ongoing speaker removal without needing past unlearned data.