A Unified Learning and Evaluation Framework for Infant Cry-based Verification.

Area of Science:

• Biometrics
• Machine Learning
• Speech Processing

Background:

• Infant crying is the primary communication method for newborns.
• Infant cry-based verification can prevent mix-ups in hospital settings.
• Current voice verification models struggle with variable-length audio and single-view evaluations.

Purpose of the Study:

• To develop a unified framework for infant cry verification.
• To improve model consistency and evaluation accuracy.
• To enhance the robustness of newborn voice verification systems.

Main Methods:

• Implemented a novel training framework using fixed-length audio segments.
• Introduced a multi-view joint evaluation strategy.
• Associated audio recordings with local segments for comprehensive analysis.

Main Results:

• Achieved consistent improvements across different verification models.
• Reduced Equal Error Rate (EER) by 10.29% for whisper-PMFA, 6.63% for X-Vector, and 5.91% for ECAPA-TDNN.
• Demonstrated enhanced model stability and evaluation accuracy.

Conclusions:

• The proposed framework significantly improves infant cry verification.
• Fixed-length segment training and multi-view evaluation enhance model performance.
• This research offers a more robust solution for newborn voice verification.