An Effective Depression Diagnostic System Using Speech Signal Analysis Through Deep Learning Methods

According to the World Health Organization (WHO), depression is one of the largest contributors to the burden of mental and psychological diseases with more than 300 million people being affected; however a huge portion of this does not receive effective diagnosis. Traditional techniques to diagnose depression were based on clinical interviews. These techniques had several limitations based on duration and variety of symptoms, due to which these methods lacked subjectivity and accuracy. Speech is tested to be an important tool in diagnosis as they carry the impression of one’s thoughts and emotions. Speech signals not only carry the linguistic feature but they also contain several other features (paralinguistic features) which can reflect the emotional state of the speaker. The analysis of these features can be used for the diagnosis of depression. With the advancement of artificial techniques and algorithms, they have become popular and are widely used in tasks of pattern recognition and signal processing. These algorithms can easily extract the features from the data and learn to recognize patterns from them. Although these algorithms can successfully recognize emotions, their efficiency is often argued. The main objective of this paper is to propose a strategy to efficiently diagnose depression from the analysis of speech signals. The analysis is performed in the following two ways: First, by considering the male and female emotions combined (gender-neutral) where they are classified into two classes, and second, separately for the male and female emotions (gender-based) for a total of four classes. Experiments conducted show the advantages and shortcomings of paralinguistic features for diagnosis of depression. During experimentation we tested several architectures by efficiently tuning the hyperparameters. For K-nearest neighbors (KNN), best attained accuracy was 86%, whereas for Multi-Layer Perceptron (MLP) architecture the accuracy attained was 87.8%. Best results were obtained from hybrid 1D-Convolutional Neural Networks (CNN) and Long Short Term Memory (LSTM) architecture with the accuracy of 88.33% and 90.07% for gender-neutral and gender-based respectively.