COMPARISON OF MACHINE LEARNING AND REINFORCEMENT LEARNING FOR DEPRESSION RECOGNITION FROM SPEECH

Reinforcement learning is one of the most promising machine learning methods used to train computer systems for decision making in order to obtain optimal results, and human speech carries rich information not only about the content of communication but also about the mental state of the speaker. Therefore, this paper investigates the possibility of applying this approach to automatically detect depression in humans based on speech data. This paper shows a comparative analysis between various traditional machine learning and reinforcement learning (RL) methods for detecting depression from human speech using speech features such as Mel-frequency cepstral coefficients. For comparison, machine learning methods were evaluated: logistic regression, support vector method (SVM), LSTM and Q-learning with reinforcement. Also, these models were tested on unbalanced, manually balanced and SMOTE (synthetic minority over-sampling technique) balanced datasets. Reinforcement learning models achieved the best F1 score value of 0.658 and AUC-ROC value of 0.628. Analysis with traditional techniques showed comparable or in some cases better performance in detecting depression in humans from speech when trained on a manually balanced dataset. The results highlight the high potential of using reinforcement learning in sequential training to develop a decision-making model for diagnosing a person's mental state from speech data.