MA-DLE: Speech-based Automatic Depression Level Estimation via Memory Augmentation

The article presents MA-DLE, a novel approach for automatic depression level estimation using speech, which enhances GRU-based models through a memory augmentation technique. This method integrates historical temporal features and dynamic memory features to improve long-range dependency capture, and employs a Hierarchical Attention Fusion (HAF) module for effective feature fusion. Evaluated on the DAIC-WOZ and E-DAIC datasets, MA-DLE achieves state-of-the-art performance, highlighting its potential for early depression detection in resource-limited settings.