Exploring Self-Supervised Models for Depressive Disorder Detection: A Study on Speech Corpora.

Area of Science:

• Computational linguistics
• Psychiatry
• Machine learning

Background:

• Automatic detection of depressive disorder from speech signals is crucial for reliable medical diagnosis.
• Limited effectiveness of deep learning models due to small depression datasets.
• Need for effective methods in low-resource speech-based depression corpora.

Purpose of the Study:

• To investigate the efficacy of speech embeddings derived from self-supervised learning (SSL) models for depression detection.
• To evaluate different SSL models (Wav2Vec 2.0, HuBERT, WavLM) and their layer depths in low-resource conditions.
• To compare SSL-based features against traditional acoustic features like MFCCs.

Main Methods:

• Extracted speech embeddings using SSL models: Wav2Vec 2.0, HuBERT, and WavLM.
• Benchmarked embeddings with traditional classifiers: SVM, Logistic Regression, Decision Tree, and Naive Bayes.
• Analyzed the impact of varying layer depths within the SSL models.

Main Results:

• Wav2Vec 2.0 and WavLM features demonstrated superior generalization compared to HuBERT features.
• WavLM features achieved a 13.1% increase in depression detection accuracy over MFCCs.
• The study identified optimal SSL models and feature extraction strategies for low-resource depression detection.

Conclusions:

• SSL models, particularly WavLM, show significant promise for enhancing depression detection accuracy in speech analysis, especially with limited data.
• The findings provide valuable insights for future research leveraging SSL for mental health applications.
• Speech embeddings offer a robust alternative to traditional features in challenging, low-resource scenarios.