Speech Depression Screening via Multi-Scale Feature Enhancement and Emotion-Aware Contrastive Learning

Area of Science:

• Computational Linguistics
• Psychiatry
• Machine Learning

Background:

• Speech-based depression screening offers a non-invasive method but faces challenges with short audio segments.
• Weak acoustic cues, limited semantic context, and overlapping participant data hinder accuracy.

Purpose of the Study:

• To develop a preliminary speech-based depression screening framework for short audio segments.
• To enhance acoustic feature extraction and representation calibration using advanced machine learning techniques.

Main Methods:

• Utilized a self-supervised Wav2Vec 2.0 model integrated with a Multi-Scale Convolution (MSC) module for acoustic feature enhancement.
• Implemented a Reference-Enhanced Contrastive Learning (ReCLR) mechanism with external emotion features for representation calibration.
• Incorporated Whisper-based transcription features to assess the utility of semantic information.

Main Results:

• Achieved an accuracy of 0.7021 and a specificity of 0.8485 for non-depressed participants in the DAIC-WOZ dataset.
• Demonstrated a sensitivity of 0.3571 for depressed participants, indicating room for improvement.
• The framework showed potential for specificity-oriented screening but requires further calibration.

Conclusions:

• Multi-scale acoustic enhancement and contrastive calibration show promise for improving specificity in short-segment screening.
• The current framework is preliminary, with limited sensitivity for depressed individuals.
• Further research is needed for threshold calibration, training stability, and external validation before clinical application.