Layer-Wise Analysis of Self-Supervised Representations for Age and Gender Classification in Children's Speech

2025-08-14

AI Summary

Age and gender classification of children’s speech is challenging due to high inter- and intra-speaker variability in pitch, articulation, and vocal tract development; however, the representational mechanisms underlying self-supervised speech models (e.g., Wav2Vec2) for speaker attributes remain poorly understood. This work systematically analyzes layer-wise speaker characteristic encoding in Wav2Vec2 variants, revealing that early layers capture acoustic individuality cues—such as age- and gender-correlated phonetic and prosodic features—while deeper layers encode linguistic content. Leveraging this insight, we propose a hierarchical feature extraction strategy tailored to children’s speech. Evaluated on PFSTAR and CMU Kids datasets, Wav2Vec2-large-lv60 with PCA-based dimensionality reduction achieves 97.14% accuracy for age classification and 98.20% for gender classification on CMU Kids. This study provides the first empirical characterization of speaker-related feature distribution across SSL model layers, establishing an adaptable modeling paradigm for child-centered perceptual interfaces.

Technology Category

• Natural Language Processing: Speech
• Machine Learning: Dimensionality Reduction/Feature Selection
• Cognitive Modeling & Cognitive Systems: Affective Computing

Application Category

• Web Mining and Content Analysis: Large pretrained models with web data
• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Children's speech presents challenges for age and gender classification due to high variability in pitch, articulation, and developmental traits. While self-supervised learning (SSL) models perform well on adult speech tasks, their ability to encode speaker traits in children remains underexplored. This paper presents a detailed layer-wise analysis of four Wav2Vec2 variants using the PFSTAR and CMU Kids datasets. Results show that early layers (1-7) capture speaker-specific cues more effectively than deeper layers, which increasingly focus on linguistic information. Applying PCA further improves classification, reducing redundancy and highlighting the most informative components. The Wav2Vec2-large-lv60 model achieves 97.14% (age) and 98.20% (gender) on CMU Kids; base-100h and large-lv60 models reach 86.05% and 95.00% on PFSTAR. These results reveal how speaker traits are structured across SSL model depth and support more targeted, adaptive strategies for child-aware speech interfaces.

Problem

• Analyzing SSL models for age and gender classification in children's speech
• Exploring layer-wise effectiveness in capturing speaker-specific traits
• Improving classification accuracy using PCA and adaptive strategies

Innovation

• Layer-wise analysis of Wav2Vec2 variants
• Early layers capture speaker-specific cues
• PCA improves classification by reducing redundancy