Deep Learning based Depression Detection using Speech Spectrograms.

Area of Science:

• Artificial Intelligence in Mental Health
• Speech Signal Processing
• Computational Psychiatry

Background:

• Depression is a widespread mental health disorder impacting daily life.
• Current diagnostic methods for depression are subjective and resource-intensive.
• Objective and scalable diagnostic tools are needed for early intervention.

Purpose of the Study:

• To explore a speech-based approach for detecting depression using deep learning.
• To evaluate the effectiveness of pre-trained convolutional neural networks (CNNs) on speech data.
• To determine if short speech segments can serve as reliable biomarkers for mental health.

Main Methods:

• Utilized the Multi-modal Open Dataset for Mental-Disorder Analysis (MODMA) with speech recordings.
• Preprocessed audio into 1-second segments and generated spectrograms.
• Fine-tuned ResNet-50, VGG-19, and EfficientNet-B0 models for classification.

Main Results:

• EfficientNet-B0 achieved the highest classification accuracy of 95.68%.
• The study demonstrated the effectiveness of transfer learning in speech-based depression detection.
• 1-second speech segments proved to be a feasible biomarker.

Conclusions:

• Deep learning models show significant potential for objective, scalable, and non-invasive depression detection.
• Speech analysis offers a promising avenue for automated mental health assessment.
• Short speech signal segments can be utilized as reliable biomarkers for mental health disorders.