Federated Primitive-Preserving Audio Transformers for Non-Identifiable Infant Cry Classification

Infant cry paralinguistics provides a non-invasive signal for early assessment of physiological and affective states. Yet, most learning approaches implicitly assume that the mapping between acoustic patterns and semantic labels is globally identifiable. In practice, this assumption often fails. Cry realizations vary substantially across infants due to physiological, developmental, and contextual factors, thereby increasing the risk of identity-related information leakage in federated aggregation. In this work, we formalize federated infant cry classification as a non-identifiable learning problem in federated settings. We show that standard federated objectives can become poorly aligned under this setting, even when average accuracy appears competitive. To address this limitation, we propose a Primitive-Preserving Audio Transformer (PPAT) that decouples shared representation learning from client-specific decision boundaries. Federation is restricted to transferable paralinguistic primitives, while semantic ambiguity is resolved locally. Across realistic federated evaluations with subject-level partitioning, the proposed approach improves worst-infant Macro F1 by up to 7 percentage points over strong federated baselines, while achieving per-class recall above 81% across all cry categories. Under the evaluated setting, the method lowers infant identity probe accuracy by more than 18%, substantially reduces observed negative transfer during training, and yields improved calibration and lower selective risk across coverage levels. Class-wise analysis reveals structured and acoustically plausible confusion patterns rather than diffuse errors. Beyond infant cry analysis, this work highlights a broader principle for decentralized learning in human-centered domains. When task semantics are not globally identifiable, robustness, stability, and reduced leakage risk may require rethinking what is shared across clients rather than how aggregation is performed.