A comparison study of widespread CNN architectures for speech emotion recognition on spectrogram

The article is devoted to a comparative study of the deep convolutional neural network (CNN) architectures recognizing basic emotions in human conversational speech using spectrograms of the speech audio signals. The paper considers a number of the most common CNN architectures: AlexNet, VGG-13, ResNet-18, MobileNet-V2, EfficientNet-B0. The research was carried out on the IEMOCAP dataset, which was labeled into four classes of basic emotions: Anger, Happiness, Neutral, Sadness. Accuracy, precision, recall, f1-score were chosen as quality metrics. A comparative analysis of recognition accuracy showed that the best results were achieved in accuracy for the simplest and least deep architectures AlexNet and VGG13 (accuracy was 0.649 and 0.662, respectively). The study also showed that with an increase in the depth and complexity of the architecture, the recognition results decrease.