How Well Do Self-Supervised Speech Models Encode Age and Gender in Children's Speech? A Layer-Wise Analysis Across Multiple Architectures

This study analyzes how self-supervised learning (SSL) models—specifically Wav2Vec2, HuBERT, Data2Vec, and WavLM—encode age and gender information in children's speech. Using layer-wise feature extraction and evaluation on PFSTAR and CMU Kids datasets, the research finds that age and gender cues are primarily captured in early to mid-level layers, with HuBERT performing best for age classification and Wav2Vec2 excelling in gender classification. These insights are crucial for practitioners developing speech processing applications, as they highlight the importance of model architecture and layer selection in capturing speaker attributes in child speech.