Machine Learning Classifiers for Voice Health Assessment under Simulated Room Acoustics.

Area of Science:

• Speech processing
• Machine learning
• Biomedical engineering

Background:

• Voice disorder detection relies on accurate vocal feature analysis.
• Machine learning (ML) models show promise but can be sensitive to data quality variations.
• Real-world acoustic conditions, like reverberation, can degrade performance.

Purpose of the Study:

• To evaluate the robustness of machine learning models for voice disorder detection under reverberation.
• To assess the impact of data augmentation with simulated reverberation on ML model performance.
• To determine optimal ML strategies for reliable voice disorder detection in varied acoustic environments.

Main Methods:

• Utilized common vocal health assessment features from steady vowel samples (135 pathological, 49 controls).
• Trained and tested six ML classifiers (including SVM, k-NN, Random Forest) on clean and reverberation-augmented data.
• Evaluated detection performance across clean, short (0.48s), and long (1.82s) reverberation conditions.

Main Results:

• Support Vector Machine (SVM) and k-Nearest Neighbors (k-NN) showed reliable accuracy with short reverberation.
• Random Forest achieved high accuracy on clean data but generalized poorly to augmented conditions.
• Reverberation significantly impacted the performance of some ML classifiers.

Conclusions:

• Training and testing ML models on reverberation-augmented data is crucial for enhancing reliability.
• ML models need to be validated under diverse acoustic conditions for practical voice disorder detection.
• Data augmentation strategies are essential for developing robust ML-based vocal health assessment tools.