Most voice anonymization research has quietly assumed the speaker is an adult.
Researchers publishing on arXiv have built and tested a voice anonymization system specifically adapted to child speech. Their pipeline draws on self-supervised learning models retrained on the MyST child speech corpus, then evaluated across single-speaker and two-speaker mixed audio. The adapted system outperformed standard anonymization on both intelligibility and perceptual quality while holding privacy protection steady. They also extended the approach to multi-speaker scenarios, combining speaker extraction with child-adapted anonymization to preserve conversational structure without exposing identities.
The gap matters because children's voices differ enough from adults - in pitch, rhythm, and acoustic properties - that models trained on adult data degrade noticeably when applied to kids. That degradation is not a minor quality issue; it is a privacy failure in any system meant to protect minors in educational recordings, therapy sessions, or child-facing apps. The researchers' results suggest domain adaptation is not optional when the subject population differs this much from the training data.
The anonymization field has spent years chasing adult benchmarks. This work is a reminder that the people who most need protection are often the ones least represented in the datasets used to build that protection.