Automatic Detection of Depression in Speech Using Ensemble Convolutional Neural Networks.

Area of Science:

• Computational linguistics
• Machine learning for healthcare
• Speech signal processing

Background:

• Depression diagnosis can be subjective and time-consuming.
• Objective, automated methods for depression detection are needed.
• Speech analysis offers a non-invasive modality for assessing mental health.

Purpose of the Study:

• To propose and evaluate a novel speech-based method for automatic depression classification.
• To leverage ensemble learning with Convolutional Neural Networks (CNNs) for improved classification accuracy.
• To benchmark the proposed system against existing methods in a standardized challenge setting.

Main Methods:

• Speech data pre-processed into log-spectrograms with balanced sampling.
• Development of a One-Dimensional Convolutional Neural Network (1D-CNN) architecture tailored for speech analysis.
• Ensemble averaging of predictions from multiple CNN models with different initializations, combined per speaker.

Main Results:

• The proposed ensemble CNN system achieved satisfactory results in the Depression Classification Sub-Challenge (DCC) at AVEC-2016.
• Performance was compared favorably against a Support Vector Machine (SVM) baseline and a single CNN classifier.
• The ensemble approach outperformed a CNN+LSTM system (DepAudionet) in the evaluated protocol.

Conclusions:

• Ensemble learning with 1D-CNNs is a promising approach for automatic depression classification from speech.
• The proposed method offers a robust and scalable solution for mental health monitoring.
• Speech-based analysis holds significant potential for objective and early detection of depression.