Depression Detection from Speech Using a Voting Ensemble Approach

Depression is a widespread mental health condition affecting millions globally, making early detection and treatment crucial for public health. This study presents a voting ensemble machine learning approach for depression detection using speech data from the Distress Analysis Interview Corpus - Wizard of Oz (DAIC-WOZ) dataset. The research methodology includes speaker diarization, data augmentation, and feature extraction using OpenSMILE's eGeMAPS v2 feature set. Synthetic Minority Over-sampling Technique (SMOTE) is applied to balance the dataset. A hard voting ensemble classifier, combining Decision Tree, Random Forest, Gradient Boosting, and K-Nearest Neigh-bor as base models, is developed and optimized through two iterations. The final model achieves competitive performance with 78.69% accuracy, 77.78% precision, 75% recall, and a 76.36% F1-score, outperforming several existing methods on the DAIC-WOZ dataset. The analysis also revealed that Shimmer (voice amplitude variation) was the most significant acoustic feature, along with spectral flux, jitter, fundamental frequency, and loudness. This automated approach contributes to early depression detection efforts to improve access to mental health care.