Multi-label Emotion Classification using modified Sequential Deep Learning Architecture for Speech Emotion Recognition

In recent years, audio classification has gained significant attention due to its wide range of applications, particularly in the field of Speech Emotion Recognition (SER). Our paper presents a unique approach to multi-label audio classification based on emotion using a modified Sequential Deep Learning architecture, applied to a combined dataset of TESS, RAVDESS, SAVEE, and CREMA-D datasets. Our model employs SoftMax to obtain multi-class probabilities, which are then used to calculate evaluation metrics and uncertainties. Additionally, we explore the performance of our model across various emotion categories, demonstrating its robustness and versatility. Our model used Dropout Layers after every Layer in the architecture to decrease the overfitting and increase the accuracy. Comprehensive experiments show that our model achieves state-of-the-art results in multi-label classification, outperforming previous methods. The method also extensively evaluates the model’s performance.