Improving speech depression detection using transfer learning with wav2vec 2.0 in low-resource environments.

Area of Science:

• Computational linguistics
• Artificial intelligence in healthcare
• Mental health informatics

Background:

• Depression is a widespread mental health disorder with significant global impact.
• Effective depression detection models are hindered by a lack of annotated speech data.
• Existing deep learning approaches require substantial labeled datasets.

Purpose of the Study:

• To develop an effective speech-based depression detection model using transfer learning.
• To address the challenge of limited annotated data in depression detection research.
• To enhance the model's ability to identify depression-related speech patterns.

Main Methods:

• Utilized transfer learning by fine-tuning wav2vec 2.0 for feature extraction.
• Integrated 1D-CNN and attention pooling for advanced segment-level feature representation.
• Employed LSTM and self-attention mechanisms to weigh depression-associated speech segments.

Main Results:

• Achieved an F1 score of 79% on the DAIC-WOZ dataset.
• Attained an F1 score of 90.53% on the CMDC dataset.
• Outperformed existing baseline models in speech-based depression detection.

Conclusions:

• The proposed transfer learning model demonstrates high efficacy in detecting depression from speech.
• The approach is particularly effective in low-resource environments with limited annotated data.
• This research offers a viable tool for scalable mental health screening and diagnosis.