MA-DLE uses memory augmentation to improve depression detection from speech
New AI model analyzes speech patterns to estimate depression levels with state-of-the-art accuracy.
The paper 'MA-DLE: Speech-based Automatic Depression Level Estimation via Memory Augmentation' (accepted at IEEE TAC) addresses a critical challenge in mental health: accurately estimating depression levels from speech. Existing deep learning approaches, primarily based on RNNs like LSTM and GRU, tend to focus on adjacent speech segments, missing long-range dependencies vital for capturing depressive symptoms.
To solve this, the authors introduce a memory-based feature augmentation method. Their 'memory bank' selectively integrates two components: (1) historical temporal features similar to the current GRU output, providing complementary context, and (2) dynamic memory features identified by variability, which capture behavioral and emotional fluctuations. A Hierarchical Attention Fusion (HAF) module then fuses these augmented features with GRU outputs. Evaluated on DAIC-WOZ and E-DAIC, MA-DLE sets new state-of-the-art results, paving the way for scalable, speech-based mental health screening.
- MA-DLE uses a memory bank that selectively integrates historical temporal features and dynamic memory features to capture long-range speech dependencies.
- A Hierarchical Attention Fusion (HAF) module fuses memory-augmented features with GRU outputs for enhanced representation.
- Achieves state-of-the-art performance on the DAIC-WOZ and E-DAIC datasets for depression level estimation.
Why It Matters
Enables scalable, non-invasive mental health screening through speech analysis, potentially improving early detection in resource-constrained settings.