Robust nasality representation learning for cleft palate-related velopharyngeal dysfunction screening in real-world...

Area of Science:

• Speech pathology
• Machine learning
• Digital health

Background:

• Velopharyngeal dysfunction (VPD) impairs speech, causing hypernasality and reduced intelligibility.
• Current screening requires specialized settings, limiting accessibility globally.
• Machine learning models struggle with real-world audio due to domain shift from varied recording conditions.

Purpose of the Study:

• To develop a robust two-stage framework for VPD screening adaptable to uncontrolled acoustic environments.
• To enhance the reliability of speech-based diagnostic tools for wider deployment.
• To mitigate performance degradation caused by domain shift in consumer devices.

Main Methods:

• A two-stage framework incorporating nasality representation pre-training using supervised contrastive learning.
• Frozen-encoder VPD screening on short audio segments with probability aggregation for decision-making.
• Evaluation using both in-domain (clinical) and out-of-domain (internet) datasets, compared against baselines like MFCC and large speech representations.

Main Results:

• Achieved perfect screening performance (macro-F1=1.000, accuracy=1.000) on in-domain clinical data.
• Demonstrated robust performance on out-of-domain internet recordings (macro-F1=0.679, accuracy=0.695), outperforming strong baselines.
• The proposed method showed significant improvements over MFCC and large pretrained speech representations on heterogeneous recordings.

Conclusions:

• Pre-training with nasality-focused representations improves robustness against recording artifacts.
• The framework supports practical deployment of VPD screening tools in real-world scenarios.
• Highlights the need for domain-robust evaluation protocols for speech-based digital health technologies.