Inter classifier comparison to detect voice pathologies.

Area of Science:

• Medical Science
• Signal Processing
• Machine Learning

Background:

• Voice pathologies arise from irregular vocal fold vibrations.
• Machine learning offers novel approaches for diagnosing voice disorders.
• Accurate detection of conditions like laryngitis, cyst, non-fluency syndrome, and dysphonia is crucial.

Purpose of the Study:

• To extract and compare features from audio signals for voice pathology detection.
• To evaluate the performance of four machine learning algorithms: SVM, Naïve Bayes, decision tree, and ensemble classifier.
• To identify the most effective combination of acoustic features for accurate diagnosis.

Main Methods:

• Utilized the SVD dataset containing audio recordings of four voice pathologies.
• Extracted a combination of 13 mel-frequency cepstral coefficients (MFCCs) along with pitch, ZCR, spectral flux, entropy, centroid, roll-off, and short-term energy.
• Applied and compared Support Vector Machine (SVM), Naïve Bayes, decision tree, and ensemble classifiers.

Main Results:

• The decision tree classifier achieved 100% accuracy, while Naïve Bayes reached 99.45%.
• SVM achieved 93.18% accuracy, and the ensemble classifier obtained 51% accuracy.
• The combined feature set significantly improved the detection accuracy of voice pathologies.

Conclusions:

• Naïve Bayes and decision tree classifiers demonstrate superior performance for voice pathology detection.
• The proposed feature combination enhances diagnostic accuracy in identifying voice disorders.
• SVM remains a commonly used algorithm for voice condition identification.