A Voice-Based Emotion Recognition System Using Deep Learning Techniques

The project aims to create an emotion recognition system based on voice using deep learning techniques. The system is based on supervised learning with artificial neural networks, enabling it to accurately predict emotions. The system's potential usage in detecting depression pathologies in the psychological area gives rise to its development. The system is designed using the KDD (Knowledge Discovery in Database) methodology and utilizes an existing database containing audio with various emotions. These audios are subjected to Multilevel Wavelet transform, decomposing the original signal into sub-signals with specific characteristics for each audio to form a training data set that is subsequently normalized, followed by the generation of the LSTM neural network architecture. Performance tests are eventually conducted on patients with depressive pathology, involving the application of the "Beck test", which indicates the severity of depression experienced by the patient. As a result, the individual reads a text that is recorded, followed by the process of feature extraction and emotion recognition performed with the pre-trained neural network. The outcome indicates that 50% of the patients exhibit severe depression, while the remainder displays milder symptoms, which is supported by the emotions detected by the system alongside the administered test.