Toward Real-World Voice Disorder Classification.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Speech Science

Background:

• Voice disorders impact daily communication and can worsen without timely diagnosis.
• Automatic home-based voice disorder classification is needed for accessibility, but faces challenges with limited resources and domain mismatch.
• Existing systems struggle with performance degradation due to noisy real-world data compared to clinical data.

Purpose of the Study:

• To develop a compact and domain-robust system for classifying voice disorders, including neoplasms and benign structural diseases.
• To address the domain mismatch between clinical and real-world noisy data for improved voice disorder detection.
• To create a system suitable for resource-constrained embedded applications.

Main Methods:

• Utilized factorized convolutional neural networks (CNNs) as a feature extractor.
• Employed domain adversarial training to extract domain-invariant features, reconciling data discrepancies.
• Focused on classifying health, neoplasm, and benign structural voice disorder categories.

Main Results:

• Achieved a 13% improvement in unweighted average recall in the noisy real-world domain.
• Maintained 80% recall in the clinical domain with minimal performance degradation.
• Reduced memory and computation usage by over 73.9%, demonstrating significant model compression.

Conclusions:

• Factorized CNNs and domain adversarial training enable derivation of domain-invariant features for voice disorder classification on limited resources.
• The system effectively eliminates domain mismatch, enhancing classification accuracy and reducing resource consumption.
• This research is the first to address model compression and noise-robustness concurrently for voice disorder classification on embedded systems.