Cross-Language Depression Detection Based on Multi-Domain Feature Alignment.

Area of Science:

• Computational linguistics
• Machine learning for mental health
• Speech signal processing

Background:

• Deep learning shows promise for detecting depression from speech.
• Speaker-specific traits and language variations create feature shifts, hindering model generalization.

Purpose of the Study:

• To develop a model that minimizes individual and language differences in speech for depression detection.
• To enhance the generalization capabilities of deep learning models for cross-individual and cross-language depression detection.

Main Methods:

• Proposed a Multi-Domain Feature Alignment (MDFA) model.
• Utilized gradient reversal techniques to reduce feature encoder sensitivity to speaker and language traits.
• Focused on depression-related patterns for improved classification.

Main Results:

• The MDFA model significantly improved generalization across individuals.
• Demonstrated enhanced performance in cross-language depression detection tasks.
• Validated on DAIC-WoZ and Androids datasets.

Conclusions:

• The MDFA model effectively addresses feature shifts caused by speaker and language differences.
• This approach leads to more robust and generalizable depression detection from speech.
• Offers a promising solution for real-world, diverse applications of mental health monitoring.