Gender-Aware Speech Emotion Recognition in Multiple Languages

This article presents a solution for Speech Emotion Recognition (SER) in multilingual setting using a hierarchical approach. The approach involves two levels, the first level identifies the gender of the speaker, while the second level predicts their emotional state. We evaluate the performance of three classifiers of increasing complexity: k-NN, transfer learning based on YAMNet, and Bidirectional Long Short-Term Memory neural networks. The models were trained, validated, and tested on a dataset that includes the big-six emotions and was collected from well-known SER datasets representing six different languages. Our results indicate that there are differences in classification accuracy when considering all data versus only female or male data, across all classifiers. Interestingly, prior knowledge of the speaker's gender can improve the overall classification performance.