An explainable voice-based age assurance system for binary age classification using acoustic features and machine learning

Reliable and transparent age assurance systems are increasingly important for online safety and access control, with voice-based approaches offering a privacy-preserving alternative to facial estimation in audio-centric environments. This study presents a novel explainable machine learning framework for binary age classification at the legally relevant threshold of <18 versus ≥18, a task that remains underexplored in prior speech-based research. Our framework integrates multiple dimensions of analysis to address key gaps in the literature: it employs a rigorous speaker-disjoint cross-validation strategy to prevent data leakage and ensure realistic performance estimates; it combines redundancy analysis, feature relevancy assessment, and SHapley Additive exPlanations (SHAP) to identify and interpret discriminative age-related acoustic cues; and incorporates trustworthiness evaluation through subgroup error, confidence, and calibration analyses to characterise model behaviour near the critical legal boundary. Using the AgeVoxCeleb dataset, we retained 8478 high-quality samples after voice activity detection and signal-to-noise ratio filtering, resulting in an imbalanced dataset (891 underage, 7587 adult samples). Among the classifiers, the baseline SVC achieved the strongest overall performance, with a macro F1 of 0.993, <18 recall of 0.98, and consistent performance across folds (SI = 0.0061). Error analysis revealed greater difficulty classifying older teens (16–17 years), while calibration analysis showed low Brier Score and ECE values on the evaluated data, though the Support Vector Machine occasionally produced high-confidence errors on minority-class samples. While our results are promising, evaluation is limited to AgeVoxCeleb, requiring external validation to confirm generalisability.