Speech-Based Emotion Recognition and PTSD Detection through Machine and Deep Learning

Abstract

This study investigates the potential of machine and deep learning algorithms for Speech Emotion Recognition (SER) and Post-Traumatic Stress Disorder (PTSD) detection through speech analysis. Traditional diagnostic methods for PTSD, which are often subjective and time-consuming, are in contrast with the automated capabilities offered by these algorithms, enabling early detection through the identification of specific speech patterns. Utilizing the RAVDESS Emotional Speech Audio dataset alongside PTSD-specific recordings, this study applies preprocessing techniques such as noise reduction and normalization to enhance the quality of the speech data. Feature extraction is performed by focusing on acoustic, linguistic, and temporal features that capture variations in the pitch, intonation, and speech rate. Both machine learning models, including Support Vector Machines (SVMs) and Random Forests, and deep learning models, such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks, have been developed and compared. Experimental results indicate that deep learning models achieve up to 91% accuracy in SER and 89% accuracy in PTSD detection, significantly outperforming traditional machine learning methods. These findings demonstrate the efficacy of multimodal integration in improving diagnostic capabilities, particularly through a combination of speech, text, and physiological data. However, the study acknowledges the limitations in generalizability across diverse populations and the practical challenges of deploying these models in real-world applications. Future work will focus on expanding the datasets to include a wider range of demographic and cultural variations, enhancing real-time monitoring capabilities, and refining model interpretability to ensure reliable performance in various contexts.