Deep Learning-Based Speech Emotion Recognition: Combining U-Net and Dense Perceptron Layers

This paper presents a hybrid model that combines U-Net and perceptron-based layers for Speech Emotion Recognition (SER). The model focuses on classifying emotions using 2D speech signal representations. It utilizes U-Net’s residual convolutional block encoder-decoder structure alongside dense perceptron layers to process Mel Frequency Cepstral Coefficients (MFCCs). The model is designed to identify emotions such as neutral, sad, and angry. When tested on the CREMA-D dataset, it achieved a training accuracy of 0.98 and a validation accuracy of 0.75. However, when evaluated on the unseen RAVDESS dataset, the model performed well only in classifying angry speech, as indicated by the confusion matrix results. Despite this outcome on unseen data, the paper highlights the potential of hybrid U-Net and perceptron models in the SER domain.