TLDR: This research explores how self-supervised learning (SSL) models like Wav2Vec2 process children’s speech for age and gender classification. It finds that early layers of these models are best at capturing speaker-specific traits, while deeper layers focus on linguistic information. Applying dimensionality reduction (PCA) further enhances accuracy and efficiency. The study demonstrates that these models can achieve high accuracy in classifying age and gender in children’s voices, offering valuable insights for developing safer and more personalized child-aware speech technologies.
Classifying age and gender from children’s speech presents unique challenges for artificial intelligence. Children’s voices exhibit high variability in pitch and articulation, along with rapid developmental changes in their vocal tracts. This makes it difficult for automated systems to accurately identify their age and gender, which is crucial for applications like personalized learning and content filtering, especially given the increasing screen time among children.
While advanced self-supervised learning (SSL) models, such as Wav2Vec2, have shown remarkable performance in tasks involving adult speech, their effectiveness in encoding speaker traits in children’s voices has remained largely unexplored. Most previous studies on children’s age and gender classification either fine-tuned these models or used complex methods to combine features from multiple layers, offering limited insight into how individual layers contribute to the classification process.
A recent study, titled Layer-Wise Analysis of Self-Supervised Representations for Age and Gender Classification in Children’s Speech, delves into this gap by providing a detailed layer-by-layer analysis of four different Wav2Vec2 models. The researchers, Abhijit Sinha, Harishankar Kumar, Mohit Joshi, Hemant Kumar Kathania, Shrikanth Narayanan, and Sudarsana Reddy Kadiri, investigated how these models process children’s speech for age and gender classification without additional fine-tuning, assessing how well the pre-trained representations generalize to children’s voices.
The study utilized two well-known children’s speech datasets: PFSTAR and CMU Kids. They extracted features from various Wav2Vec2 model variants, including base-100h, base-960h, large-960h-lv60, and large-960h-lv60-self. These features were then fed into a simple Convolutional Neural Network (CNN) classifier. To further refine the process and enhance performance, Principal Component Analysis (PCA) was applied for dimensionality reduction, aiming to isolate and retain only the most critical features.
Key Findings from the Layer-Wise Analysis
The research revealed significant insights into how different layers of SSL models encode age and gender cues. The findings indicate that the early layers (typically layers 1 to 7) are most effective at capturing speaker-specific attributes, such as age and gender. These layers primarily focus on fundamental, low-level acoustic features. In contrast, deeper layers progressively abstract these details, concentrating more on linguistic and phonetic representations, making them less effective for direct age and gender classification.
The study also demonstrated that applying PCA for dimensionality reduction consistently preserved or even improved classification performance. This suggests that the speaker-relevant information is highly concentrated within a smaller subset of principal components. By reducing the feature dimensions, the models not only became more computationally efficient but also enhanced their generalization capabilities by removing redundant or noisy components.
Performance Highlights
The Wav2Vec2 models significantly outperformed traditional methods using Mel-Frequency Cepstral Coefficients (MFCCs) for age classification across both datasets. For gender classification, improvements were also observed, particularly with larger model variants. On the CMU Kids dataset, the Wav2Vec2-large-lv60 model achieved exceptional results, with 97.14% accuracy for age classification and 98.20% for gender classification after applying PCA. On the PFSTAR dataset, the base-100h model achieved 86.05% accuracy for age classification, and the large-960h-lv60-self model reached 95.00% accuracy for gender classification with reduced feature dimensions.
Also Read:
- Transparent AI for Interpreting Assessment in Higher Education
- Enhancing Speech Clarity with an Alternating Approach-Putt Model
Implications for Child-Aware Speech Systems
These results offer valuable insights into the internal workings of SSL models and how speaker traits are distributed across their depth. The discovery that early layers are crucial for age and gender identification, combined with the effectiveness of dimensionality reduction, supports the development of more targeted and adaptive strategies for child-centric speech interfaces. This research paves the way for creating more robust and efficient systems that can reliably inform content filtering algorithms and ensure safer, more appropriate online experiences for children.
