Infant Cry Intelligence Programme

Problem

Infant cry is a primary communication channel in early life, yet interpretation remains informal, delayed and highly dependent on caregiver experience. In settings with limited neonatal expertise, that gap has clinical consequences.

Context

This programme spans four scholarly projects from the research record: adaptive bandit CRNN modality selection, a domain-agnostic causal-aware audio transformer, privacy-enhancing federated transformers with denoising regularisation, and an autonomous surveillance Android application for artificial parenting support.

Dataset

Evaluations draw on established cry corpora including Baby Chillanto and Donate-a-Cry, with environmental noise overlays used to stress domain generalisation. Additional in-situ collections are documented in the corresponding papers.

Methodology

The work combines attentive convolutional-recurrent models with multi-armed bandit modality selection, denoising autoencoders, federated training, causal-aware hierarchical audio transformers, and primitive-preserving federation. Cry semantics are treated as non-identifiable across infants, so privacy and local decision boundaries are first-class research problems.

Model / architecture

Architectures include an attentive CRNN with bandit modality selection (Complex & Intelligent Systems, 2025); DACH-TIC, a domain-agnostic causal-aware hierarchical audio transformer; federated transformers with denoising regularisation; and a CNN-based audio cry classifier deployed in an Android parenting application.

Results

Published results include gains in accuracy and macro-F1 over strong audio-transformer baselines, reduced domain gap under unseen acoustic conditions, improved worst-infant macro-F1 in federated evaluations, and suppressed identity-probe leakage relative to standard federated objectives. The Android application operationalises autonomous cry surveillance for caregivers.

Evaluation metrics

Scholarly projects

4 linked publications

Deployment

Android artificial parenting app

Privacy posture

Federated / non-identifiable design

Challenges

Subject-level variability, background noise, small and imbalanced clinical datasets, identity leakage in federated aggregation, and the engineering gap between laboratory models and a mobile app caregivers can use daily.

Impact

A path toward real-time distress signalling for caregivers and clinicians, particularly where neonatal specialist coverage is thin. Supported by a Google PhD Fellowship in speech processing and health analytics.

Tools / technology

• Python
• PyTorch
• Audio transformers
• Federated learning
• Android
• Signal processing