Multi-Task Learning for Audio-Based Infant Cry Detection and Reasoning.

Area of Science:

• Machine Learning
• Infant Health Monitoring
• Signal Processing

Background:

• Infant cry analysis is vital for understanding infant well-being, but limited datasets and individual voice variations hinder model performance.
• Existing models struggle with generalization due to data scarcity and subject-specific acoustic differences in infant cries.

Purpose of the Study:

• To develop a robust multi-task model for Infant Cry Detection and Reasoning (ICDR) that overcomes data limitations and subject variability.
• To improve the accuracy and generalization capabilities of AI models in interpreting infant vocalizations.

Main Methods:

• Proposed a multi-task learning framework leveraging two datasets to increase data diversity.
• Introduced an efficient attention module for inter-task feature enrichment.
• Implemented an intra-task contrastive mixture of experts (CMoE) module to reduce subject variance and enhance representation consistency.

Main Results:

• The ICDR model demonstrated superior performance compared to state-of-the-art methods in infant cry detection and reasoning.
• Achieved significant improvements in F1-score (2-9%) across extensive cross-subject experiments.
• Validated the effectiveness of multi-task learning and the CMoE module in enhancing model generalization.

Conclusions:

• Multi-task learning with inter-task attention and intra-task CMoE significantly boosts the generalization ability of infant cry analysis models.
• The proposed ICDR model offers a promising solution for more reliable infant cry interpretation in real-world applications.