⚡ electrical engineering

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

This paper proposes a child-centric voice anonymization framework using domain-adapted self-supervised learning models that effectively preserves speech intelligibility and quality while ensuring strong privacy protection in both single-speaker and multi-speaker scenarios.

Original authors: Pranav Tushar, Xiao Xiao Miao, Rong Tong

Published 2026-06-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Pranav Tushar, Xiao Xiao Miao, Rong Tong

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a recording of a child talking. You want to protect their identity (so no one knows who they are) but keep their words clear (so people can still understand what they are saying). This is called voice anonymization.

Most current systems for doing this are like "adult-sized gloves." They are designed and trained on adult voices. When you try to put them on a child's voice, they don't fit right. The child's voice sounds higher-pitched, more variable, and sometimes a bit "wobbly" (like a child learning to speak), which confuses the adult-trained system. The result? The child's words become garbled, or the system accidentally changes the child's voice to sound like an adult, which ruins the natural feel.

This paper introduces a new approach: Child-Centric Voice Anonymization. Think of it as tailoring a custom suit specifically for a child, rather than trying to shrink an adult suit down.

Here is how they did it, broken down into simple steps:

1. The "Deconstruct and Rebuild" Strategy

The researchers used a smart system (based on Self-Supervised Learning, or SSL) that acts like a high-tech chef.

The Ingredients: When a child speaks, the system breaks the voice down into three separate "ingredients":
1. The Recipe (Content): The actual words and meaning.
2. The Seasoning (Prosody): The rhythm, emotion, and pitch (how high or low the voice goes).
3. The Chef's Signature (Speaker Identity): The unique "fingerprint" that tells you who is speaking.
The Swap: To protect privacy, the system keeps the "Recipe" and "Seasoning" exactly as they are. It throws away the original "Chef's Signature" and replaces it with a fake one from a pool of other voices.
The Rebuild: It mixes these back together to create a new voice that says the same words with the same rhythm, but sounds like a different person.

2. The "Child-Specific" Upgrade

The problem with the old system was that the "Chef" (the computer model) had only ever cooked with adult ingredients. When given child ingredients, it messed up.

The Fix: The researchers took the computer models responsible for understanding the words and rebuilding the voice and re-trained them specifically on children's voices (using a dataset called MyST).
The Result: Now, the system understands that a child's voice is naturally higher and more variable. It doesn't try to force the child to sound like an adult. Instead, it swaps the child's identity with another child's identity (or an AI-generated child voice), keeping the "child-like" feel intact.

3. The "Two-Person Conversation" Challenge

Real life isn't just one child talking; it's often a child talking to a teacher or another child. The researchers tested what happens when two people talk at once (a "mixture").

The Process: They built a pipeline that first acts like a spotlight, shining only on the child they want to protect (Target Speaker Extraction). Once the spotlight isolates the child, the anonymization "chef" swaps their identity. Then, the spotlight fades, and the two voices are mixed back together.
The Finding:
- Privacy: The system was very good at hiding the child's identity, even when they were talking over someone else.
- Clarity: The system worked well when the child talked to an adult. However, when two children talked over each other, it became much harder for the computer to separate them. This made the final words a bit harder to understand. The paper notes that the "spotlight" struggles to tell two similar-sounding kids apart, which causes the confusion, not the anonymization itself.

4. What They Found (The Bottom Line)

Better Fit: By training the system on children's data, the anonymized voices sounded much more natural and intelligible than systems trained on adults.
Keeping the "Child" Vibe: The system successfully hid who the child was without making them sound like an adult. Listeners still heard a child speaking, just a different one.
The Limitation: The biggest hurdle remains separating two children talking at the same time. If the computer can't clearly separate the voices to begin with, the privacy protection works, but the words get a bit muddled.

Summary Analogy

Imagine you have a group of kids playing in a park, and you want to blur their faces in a video so they can't be recognized, but you still want to hear their laughter and words clearly.

Old Method: You used a blur filter designed for adults. It made the kids' faces look weird and distorted their voices to sound deep and grown-up.
New Method: You used a filter trained specifically on kids. It blurred their faces perfectly while keeping their high-pitched giggles and clear words.
The Catch: If two kids are standing right next to each other, the camera has a hard time telling them apart, so the blur gets a little messy in that specific spot.

The paper concludes that to protect children's privacy in the future, we must build tools that understand children's voices, rather than forcing children's voices into adult-shaped tools.

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

Problem Statement
Current voice anonymization systems are predominantly developed and evaluated on adult speech. When applied to children's speech, these systems suffer from significant performance degradation. Children's speech constitutes a distinct acoustic domain characterized by higher fundamental frequencies, greater articulatory and prosodic variability, and developmental disfluencies. This creates a substantial mismatch for pipelines trained on adult data, leading to degraded intelligibility and perceptual quality. Furthermore, existing approaches often convert children's voices into adult-like voices, distorting age-dependent acoustic cues essential for child-centered applications. Finally, prior work largely ignores multi-speaker scenarios, which are common in real-world child interactions (e.g., classrooms), where selective anonymization of a target child while preserving conversational structure is required.

Methodology
The paper proposes a child-centric anonymization pipeline based on Self-Supervised Learning (SSL) that is adapted specifically for the child speech domain. The system operates in three main stages:

Disentanglement: An input waveform is decomposed into three representations:
- Content: Soft content features extracted via a HuBERT-based encoder.
- Prosody: Pitch contour ( $f_0$ ) extracted via a pitch extractor.
- Speaker: Speaker embeddings extracted via ECAPA-TDNN.
Anonymization: The source speaker embedding is replaced by a reference embedding drawn from a controlled child speaker pool. This pool consists of AI-generated voices (Typecast, SpeechGen) screened for naturalness and age consistency, ensuring the anonymized output retains child-like acoustic characteristics (child-to-child anonymization).
Resynthesis: The anonymized speech is reconstructed using a HiFi-GAN vocoder, feeding the unchanged content and prosody representations alongside the anonymized speaker embedding.

Domain Adaptation Strategy
To address the acoustic mismatch between adult-trained models and child speech, the authors fine-tune specific components of the pipeline on the MyST corpus (a child speech dataset):

HuBERT Content Encoder: Fine-tuned to improve the extraction of linguistically meaningful representations from child utterances, which differ in pitch range and formant frequencies.
HiFi-GAN Vocoder: Fine-tuned to prevent the resynthesis of child speech from shifting toward adult-like acoustics, thereby preserving child-specific spectral and prosodic characteristics.
Speaker Encoder (ECAPA-TDNN): Kept fixed, as the anonymization identity is determined by the replacement of the source embedding with a reference from the child pool.

Multi-Speaker Extension
The system is extended to two-speaker mixtures (Target Speaker Anonymization) by integrating a Target Speaker Extraction (TSE) module. The pipeline:

Uses a Conformer-based TSE model to isolate the target speaker's signal from the mixture, conditioned on a reference utterance.
Anonymizes the extracted target signal using the child-adapted SSL pipeline.
Reconstructs the mixture by replacing the original target component with the anonymized version while leaving the non-target speaker unmodified.

Key Contributions

Systematic Child-Domain Adaptation: A comprehensive study adapting SSL-based anonymization pipelines for child-to-child voice anonymization, evaluated on in-domain (MyST) and zero-shot cross-accent benchmarks (MPS, SpeechOcean).
Multi-Speaker Evaluation: An extension to two-speaker mixtures across three age-group pairings (Adult-Adult, Child-Adult, Child-Child), demonstrating the feasibility of selective anonymization in conversational settings.
Preservation of Child Characteristics: The introduction of a controlled child speaker pool and domain-adapted vocoders to ensure anonymized speech remains perceptually child-like, avoiding the distortion of age-dependent cues found in adult-oriented systems.

Experimental Results

Single-Speaker Performance:
- Privacy: The fully adapted system (FT/FT) achieved the strongest privacy protection, with Equal Error Rates (EER) reaching up to 45.09% on the MyST test set, significantly higher than the original system (43.80%) and the signal-processing baseline (B2).
- Intelligibility: The adapted system reduced Word Error Rates (WER) compared to the non-adapted SSL baseline (16.64% vs. 17.31% on MyST), demonstrating that domain adaptation preserves speech utility.
- Perceptual Quality: Human evaluations indicated that the child-adapted system (SSL-FT) received higher naturalness and fluency ratings than the B2 baseline and more consistently preserved the perception of the speaker as a child while maintaining low speaker similarity scores.
Multi-Speaker Performance:
- Privacy Stability: Privacy protection (EER) remained robust and relatively stable across varying overlap ratios (0–100%) and age pairings.
- Utility Constraints: Intelligibility (measured by target WER) degraded as overlap increased. This degradation was most severe in Child-Child (CC) mixtures, where acoustic similarity between speakers made target extraction difficult. Diarization Error Rates (DER) also increased significantly for CC pairs.
- Decoupling: The results highlight a decoupling between privacy and utility in multi-speaker scenarios: privacy remains stable, but utility is primarily bounded by the quality of target speaker extraction, particularly when extracting a child from a mixture containing another child.

Significance and Limitations
The paper claims that child-specific adaptation is critical for practical speech anonymization systems involving minors. By preserving perceived childness while obfuscating identity, the proposed method addresses the limitations of adult-centric systems that distort age-related cues.

However, the authors note several limitations:

Evaluation Mismatch: Many evaluation metrics (ASR, diarization, speaker verification, and perceptual quality predictors) remain trained on adult data, which may introduce domain mismatches when evaluating child speech.
Extraction Challenges: The performance in Child-Child mixtures is limited by the adult-trained Target Speaker Extraction model, suggesting that future improvements in child-robust extraction are necessary to enhance utility without compromising privacy.
Age Quantification: Automatic age prediction models generalized poorly, necessitating reliance on human perceptual judgments for age consistency assessment.

In conclusion, the work demonstrates that adapting SSL-based pipelines to the child domain effectively balances privacy, intelligibility, and the preservation of child-specific acoustic characteristics, while identifying target speaker extraction as the primary bottleneck for multi-speaker child anonymization.