BSER-Bengali Speech Emotion Recognition Based on SUBESCO Dataset

Speech Emotion Recognition (SER) plays a significant role in improving human-computer interaction (HCI) by allowing systems to understand and respond to users’ emotional states. This paper uses advanced deep learning techniques to present a Bengali Speech Emotion Recognition (BSER) system based on the SUBESCO dataset. The dataset comprises $\mathbf{7, 0 0 0}$ balanced audio recordings across seven emotions: anger, disgust, fear, happiness, neutral, sad, and surprise. To enhance the dataset, we applied data augmentation techniques, expanding it to $\mathbf{2 8, 0 0 0}$ samples for robust model training. Three machine learning models and one deep learning model were evaluated, including Logistic Regression (LR), Random Forest (RF), k-nearest Neighbor (KNN), and a 1D Convolutional Neural Network (CNN). The CNN model outclassed others, achieving an accuracy of 97.66% by effectively capturing spatial hierarchies in speech features such as Mel Frequency Cepstral Coefficients (MFCC), ZeroCrossing Rate (ZCR), and Log Mel-Spectrogram. The results highlight the potential of CNN architectures for accurate emotion recognition in Bengali speech. Future work aims to improve the model by exploring hybrid architectures like CNN-LSTM, incorporating multi-modal data, and expanding the dataset to improve performance in real-world conditions. The BSER system paves the way for enhanced HCI, automated customer service, and emotion-aware technologies in the Bengali-speaking world.