A cross-linguistic depression detection method based on speech data.

Area of Science:

• Artificial Intelligence
• Computational Linguistics
• Psychiatry

Background:

• Depression is a widespread mental health issue impacting individuals and their communities.
• Automated depression detection using speech data is an emerging field within artificial intelligence.
• Current AI models for depression detection often lack cross-linguistic applicability due to reliance on monolingual data.

Purpose of the Study:

• To develop a transfer learning method for cross-linguistic depression detection from speech.
• To address the limitation of monolingual data in existing AI models for depression detection.
• To enhance the generalization capability of speech-based depression detection models across different languages.

Main Methods:

• Proposed the Deep Covariance Alignment Network (DCAN), a transfer learning approach.
• Transferred models trained on English speech data (source) to Chinese speech data (target).
• Utilized down-sampled speech data (1 kHz), Convolutional AutoEncoder features, and manually selected features for analysis.

Main Results:

• Achieved 88.7% accuracy on English and 81.1% on Chinese datasets.
• Outperformed models trained solely on English data by an average of 21.9%.
• Demonstrated a 4% improvement over other mainstream transfer learning methods, indicating cross-linguistic commonalities in speech and depression.

Conclusions:

• The proposed DCAN method enhances the generalization of speech-based depression detection across linguistic domains.
• This approach reduces the need for extensive language-specific speech datasets.
• The findings suggest the method's utility in supporting depression diagnosis and highlight the potential for broader language inclusion in future research.