Application of Pre-trained Model-based Speech Analysis in Depression Detection

Detecting depression at an early stage is critical for both public health and the well-being of patients. Even though there has been much improvement with the use of automatic depression assessment technologies based on machine learning, a number of problems still exist: demographic confounding factors; complicated feature engineering; and data privacy worries; restricted sample size; and so on. To address these challenges, this study proposes a depression detection method based on pre-trained models that utilize speech data from the MODMA dataset. The innovation of this study lies in the use of pre-trained models to solve problems related to small sample sizes, while adapting the model to the Chinese language contexts and enhancing its generalization capabilities. At the same time, this study systematically examines the effectiveness of verbal tasks associated with different emotions and categories in detecting depression. These findings not only improve the accuracy, practicality, and privacy protection of depression diagnosis but also offer new insights for achieving more personalized and precise mental health management. Future research will further explore the model's generalization ability across diverse datasets and aim to apply it towards developing practical applications.