Domain-Agnostic Causal-Aware Audio Transformer for Infant Cry Classification

TL;DR

This paper introduces DACH-TIC, a novel causal-aware audio transformer that improves infant cry classification accuracy and robustness across different environments by integrating causal attention, hierarchical learning, and domain generalization.

Contribution

The paper presents a domain-agnostic, causal-aware hierarchical transformer model for infant cry classification, enhancing robustness and interpretability over existing methods.

Findings

Outperforms state-of-the-art baselines in accuracy and macro-F1 score.

Achieves only 2.4% performance gap on unseen environments.

Demonstrates improved causal fidelity and domain generalization.

Abstract

Accurate and interpretable classification of infant cry paralinguistics is essential for early detection of neonatal distress and clinical decision support. However, many existing deep learning methods rely on correlation-driven acoustic representations, which makes them vulnerable to noise, spurious cues, and domain shifts across recording environments. We propose DACH-TIC, a Domain-Agnostic Causal-Aware Hierarchical Audio Transformer for robust infant cry classification. The model integrates causal attention, hierarchical representation learning, multi-task supervision, and adversarial domain generalization within a unified framework. DACH-TIC employs a structured transformer backbone with local token-level and global semantic encoders, augmented by causal attention masking and controlled perturbation training to approximate counterfactual acoustic variations. A domain-adversarial objective promotes environment-invariant representations, while multi-task learning jointly optimizes cry type recognition, distress intensity estimation, and causal relevance prediction. The model is evaluated on the Baby Chillanto and Donate-a-Cry datasets, with ESC-50 environmental noise overlays for domain augmentation. Experimental results show that DACH-TIC outperforms state-of-the-art baselines, including HTS-AT and SE-ResNet Transformer, achieving improvements of 2.6 percent in accuracy and 2.2 points in macro-F1 score, alongside enhanced causal fidelity. The model generalizes effectively to unseen acoustic environments, with a domain performance gap of only 2.4 percent, demonstrating its suitability for real-world neonatal acoustic monitoring systems.