Voice-Based Depression Classification Using Mel-Spectrogram Representation and Texture-Color Feature Extraction

Depression is a mental health disorder that affects millions of people worldwide and often goes undetected in its early stages due to the limitations of diagnostic methods that still rely on clinical observation and subjective interviews. Delays in the early detection of depression can worsen the condition of sufferers and increase the risk of extreme actions such as suicide. To overcome this, an efficient and accurate computing system is needed. This study aims to evaluate the effectiveness of the Visual Feature Engineering approach in classifying depression based on texture and color features extracted from the Mel-Spectrogram representation Mel-Spectrogram representation, which is then tested using the DAIC-WOZ dataset. The Mel-Spectrogram representation of voice recordings is processed using manual feature extraction, including color (RGB) and texture (LBP, GLCM) feature extraction. Classification performance was tested using Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Decision Tree (DT) models. The optimal configuration was obtained from a combination of RGB + GLCM features with the SVM model. This configuration produced the best evaluation metrics on the test dataset with an accuracy of 0.89, recall of 0.95, precision of 0.84, and F1-Score of 0.89. The Visual Feature Engineering approach using a combination of RGB and GLCM features from the Mel-Spectrogram with the SVM model proved to be an effective and efficient method for voice-based depression classification.