Objective:

To improve the robustness of speech-based machine learning models for screening velopharyngeal dysfunction (VPD) in uncontrolled acoustic environments, addressing challenges posed by domain shifts.

Approach:

• Nasality Representation Pre-training: Utilizes supervised contrastive learning (SupCon) with an auxiliary dataset to form oral-context versus nasal-context supervision, enhancing the model's ability to distinguish between different speech contexts.
• Frozen-encoder VPD Screening: Employs lightweight classifiers on 0.5-second audio chunks with probability aggregation for recording-level decisions, ensuring efficient processing in real-world applications.

Key Findings:

• The proposed method achieved ceiling performance in standardized clinical conditions (macro-F1 = 1.000, accuracy = 1.000), indicating its effectiveness in controlled settings.
• In-domain nested 5-fold cross-validation yielded a mean performance of macro-F1 = 0.981 ± 0.022 and accuracy = 0.985 ± 0.016, demonstrating strong reliability across varied samples.
• On out-of-domain recordings, the proposed method outperformed previous baselines with macro-F1 = 0.679 and accuracy = 0.695, highlighting its potential for real-world application.

Interpretation:

Learning a nasality-focused representation enhances model robustness against recording artifacts, facilitating better performance when transitioning from controlled to real-world environments.

Limitations:

• The study primarily focuses on specific datasets and may not generalize to all real-world scenarios, particularly in diverse acoustic environments.
• Performance may vary with different consumer devices and recording conditions, which could impact the reliability of the model in practice.

Conclusion:

The approach supports the development of VPD screening tools, though further validation is needed to confirm its effectiveness in diverse real-world settings.