Real-Time Speech Signal Classification for Emotion Detection Using Deep Belief Convolutional Neural Network with Multi-Head Self-Attention

In recent years, the recognition of human emotions from speech signals has developed a significant attention, determined by the rising demand for affective human and computer interaction. However, the emotion detection from speech signals is a challenging task; hence the existing Convolutional Neural Network (CNN) offered a promising solution for speech emotion recognition. But, it has led to the need of low-latency processing, variability, robust to noise, and efficient computation. Hence, this research aims to develop a real-time speech signal classification system for emotion recognition using Deep Belief Convolutional Neural Network (DBCNN) with Multi-Head Self-Attention with Contextual Fusion Mechanism (MHSA-CFM). This process begins with the collection of input signals and preprocessed with Neural Stochastic Differential Equations (NSDE), as it utilizes stochastic differential equations to improve the flexibility. Then, 3D Log-Mel Spectrogram is utilized for feature extraction, where it represents audio signals as 3D visual representation of time, frequency, and amplitude. Finally, DBCNN-MHSA-CFM is employed to classify the emotions according to speech signals. From the results, the proposed DBCNN-MHSA-CFM achieved outperforming results in terms of accuracy with 93.24% when compared to the existing Gaussian Mixture Modeling (GMM) model.