Automatic Depression Detection Using Smartphone-Based Text-Dependent Speech Signals: Deep Convolutional Neural...

Area of Science:

• Speech analysis
• Computational linguistics
• Mental health technology

Background:

• Automatic depression diagnosis via speech analysis is a promising area for mental health support.
• Previous research focused on acoustic properties, but large-scale studies on non-English speakers are limited.
• Identifying depression in non-English speakers using acoustic features requires further investigation.

Purpose of the Study:

• To propose a framework for automatic depression detection using large-scale acoustic characteristics in the Korean language.
• To evaluate the effectiveness of a deep convolutional neural network (CNN) model for depression detection.
• To compare the proposed CNN model with conventional machine learning and pretrained models.

Main Methods:

• Recruited 153 major depressive disorder patients and 165 healthy controls.
• Recorded participants' speech while reading predefined sentences on smartphones.
• Evaluated three approaches: conventional machine learning, a proposed CNN model, and pretrained networks using log-Mel spectrograms.

Main Results:

• The proposed CNN model achieved the highest accuracy of 78.14% in automatically detecting depression from speech.
• Deep-learned acoustic characteristics demonstrated superior performance compared to conventional and pretrained models.
• The study successfully identified depression using text-dependent read speech tasks.

Conclusions:

• Speech analysis of text-dependent reading tasks can aid in the automatic prediction of depression status.
• The smartphone-based method is accessible and can contribute to objective depression identification.
• This approach offers a potential tool for monitoring and identifying depressive states.