Detection of Voice Disorder Using Spectral and Statistical Audio Features and SVM-RBF

The detection of voice disorders using voice signal processing and machine learning has emerged as a promising early diagnosis and treatment approach. This paper presents a machine learning-based automatic voice disorder detection framework extracting and selecting optimum features from voice audio signals. The practical challenge of class-imbalanced scenario of disorder and healthy voices data is addressed using the Synthetic Minority Over-sampling Technique (SMOTE). Several features, i.e., Mel-Frequency Cepstral Coefficients (MFCC), Chroma Short-Time Fourier Transform (STFT), zero crossing rate, root mean square, skewness, and kurtosis; have been studied in spectral and statistical domains. Finally, the optimum features are selected using the Recursive Feature Elimination with Cross-Validation (RFECV) technique to reduce the classifier model's complexity. A support vector machine (SVM) classifier with a Radial basis function kernel is developed, prioritizing simplicity and efficiency. The SVM model is trained and tested using the well-accepted VOICED database of PhysioNet Data Bank where the audio is captured using a smartphone. The performance of the overall framework shows an accuracy of 95% and 100% precision and specificity, demonstrating its resilience for precise evaluations and smartphone-based diagnosis applications.