تقنيات تحليل الصوت: نحو تشخيص غير جراحي لاضطرابات الغدة الدرقية

Abstract

This study aims to develop a non-invasive approach for diagnosing thyroid disorders through voice analysis using signal processing techniques and spectral feature extraction. This method could contribute to early detection and reduce the need for traditional medical examinations. The Saarbruecken Voice Database was used, which contains voice recordings of 1,002 speakers with voice disorders and 851 healthy speakers. A subset of 173 recordings from patients and 53 recordings from healthy individuals was selected for analysis.

Five key spectral features were extracted using Python, specifically the Librosa, PySoundFile, and NumPy libraries: Spectral Centroid, Spectral Flux, Spectral Spread, Spectral Rolloff, Spectral Kurtosis. The Spectral Gravity Center (SGC) was calculated for each group, and the Euclidean distance was used to determine the proximity of new samples to the two groups (healthy and affected). This approach enabled the construction of a simple yet effective classification model. The model achieved a classification accuracy of 90%, demonstrating strong performance in distinguishing between healthy and affected voices. The spectral features showed clear differences between the two categories, reinforcing their potential as a non-invasive diagnostic tool. However, some misclassifications were observed, suggesting the possibility of improving the model using advanced machine learning techniques.

This study confirms that voice analysis can be an effective method for non-invasive thyroid disorder diagnosis. It is recommended to enhance the model using more advanced classification techniques such as neural networks or Support Vector Machines (SVM) and to increase the dataset size for more accurate results. Additionally, developing an intelligent application for automated voice analysis could assist doctors in early diagnosis.