INFANT CRYING DETECTION IN REAL-WORLD ENVIRONMENTS.

Area of Science:

• Machine Learning
• Audio Signal Processing
• Infant Health Monitoring

Background:

• Existing infant cry detection models often lack generalizability due to training on controlled, in-lab data.
• The performance of these models in noisy, real-world environments remains largely unexamined.

Purpose of the Study:

• To evaluate the external validity of established machine learning approaches for infant cry detection in everyday settings.
• To develop and validate a novel model and dataset for improved real-world cry detection.

Main Methods:

• Collected and annotated over 780 hours of real-world infant audio data from home environments.
• Evaluated several machine learning models, including one combining deep spectrum and acoustic features.
• Tested model performance on both in-lab and real-world datasets.

Main Results:

• A model leveraging deep spectrum and acoustic features achieved an F1 score of 0.613 in real-world settings.
• Models trained on in-lab data significantly underperformed when tested on real-world data (F1: 0.236 vs. 0.656).
• The novel dataset and model demonstrate enhanced external validity for infant cry detection.

Conclusions:

• Infant cry detection models trained in controlled settings exhibit poor performance in real-world environments.
• The newly collected dataset and proposed model are crucial for advancing accurate, real-world infant cry detection.
• Further research is needed to improve the robustness of cry detection systems for practical applications.