Speech Emotion Recognition Using 1D/2D Convolutional Neural Networks

Over the last few decades, emotion recognition has been a hot topic of research in the affective computing community. The automatic identification of emotions from raw speech signals is highly challenging and depends on multiple factors, including: the utterance length, the speaker language, gender or accent. In addition, the process is highly subjective because people can perceive emotions differently. In that regard, the goal of this paper is to evaluate some state-of-the-art deep convolutional neural networks (DCNNs) architectures receiving as input various ID/2D speech feature representations, conduct experiments on a publicly available dataset (Ryerson Audio-Visual Database of Emotional Speech and Song Dataset - RA VDESS) and identify which architecture has the best performance in the discrete emotion classification task.