SVM-based identification of pathological voices.

Area of Science:

• Speech and Hearing Sciences
• Biomedical Engineering
• Machine Learning in Healthcare

Background:

• Dysphonia encompasses various voice disorders affecting vocal fold function.
• Accurate identification of pathological voices is crucial for diagnosis and treatment.
• Acoustic analysis offers objective measures for voice assessment.

Purpose of the Study:

• To develop and evaluate a support vector machine (SVM) based classification method for identifying pathological voices.
• To assess the effectiveness of SVM with different kernels in distinguishing between healthy and dysphonic voices.
• To explore the utility of principal component analysis (PCA) in feature reduction for voice data.

Main Methods:

• Collected voice recordings of the vowel "a" from 214 subjects (181 pathological, 33 healthy).
• Extracted 25 acoustic parameters from each voice sample.
• Applied principal component analysis (PCA) for dimensionality reduction.
• Trained and evaluated a soft-margin support vector machine (SVM) with three kernel types.

Main Results:

• The SVM-based approach demonstrated promising effectiveness in classifying pathological voices.
• Different SVM kernel combinations were investigated, yielding varying performance metrics.
• PCA successfully transformed the acoustic dataset into a new, potentially more informative feature space.

Conclusions:

• The proposed SVM method is a potentially effective tool for pathological voice identification.
• Further investigation into parameter tuning and kernel selection can optimize performance.
• This approach contributes to objective voice disorder assessment through machine learning.