Child-Centric Voice Anonymization in Single and Multi-Speaker Speech via Domain-Adapted SSL Models

Voice anonymization is essential for protecting the identity of speakers in sensitive recordings, yet most existing systems are trained on adult speech. This creates a significant acoustic mismatch when applied to children, whose speech is characterized by higher pitch, distinct formant frequencies, and greater prosodic variability. This paper introduces a child-centric anonymization pipeline that adapts self-supervised learning (SSL) models to the child speech domain, specifically targeting the preservation of linguistic content and natural child-like vocal characteristics.

The researchers utilized an SSL-based anonymization framework that disentangles speech into content, prosody, and speaker identity representations. To make this system child-aware, they fine-tuned the HuBERT content encoder and the HiFi-GAN vocoder using the MyST child speech corpus. Furthermore, they replaced the standard adult speaker pool with a curated set of synthetic child-like voices. For multi-speaker scenarios, the team integrated a Conformer-based target speaker extraction (TSE) model to isolate the target child's voice before applying the anonymization pipeline.

Experimental results demonstrate that full domain adaptation (fine-tuning both the encoder and vocoder) provides the best balance between privacy and utility. Compared to adult-oriented baselines, the child-adapted system significantly improves intelligibility and perceptual quality while successfully concealing the speaker's identity. In multi-speaker settings, the researchers observed a decoupling between privacy and utility: while privacy protection remains robust regardless of speaker overlap, intelligibility is primarily constrained by the performance of the target speaker extraction model. This effect is most pronounced in child-child mixtures, where acoustic similarity makes source separation particularly challenging.