Speech-based Depression Estimation

Abstract

This thesis focuses on using machine learning to develop objective methods for esti- mating depression, thus addressing the limitations in current diagnostic practices. The research introduces a novel pipeline for extracting audio features and text embeddings from the DAIC-WOZ dataset. Specifically the PyAudioAnalysis library was utilized for audio feature extraction and GloVe embeddings for text features. A role-based extrac- tion method was implemented to independently process features for the participant and the interviewer, providing insights into the significance of each role in depression esti- mation and the influence of interaction dynamics on predictive accuracy. In this study machine learning techniques are applied such as Support Vector Machines (SVM) and XGBoost models, to improve depression detection. The primary goal is to identify the most effective combination of features and algorithms that can enhance the accuracy and reliability of depression prediction models. Key findings indicate that text-based features, particularly GloVe embeddings, outperform traditional audio features, achiev- ing an AUC score of 0.74 for text-based models compared to 0.66 for audio-based models. The study also explores balancing techniques, noting that while SMOTE im- proved model performance, the choice of features remains critical.