A Comparative Study on Different Labelling Schemes and Cross-Corpus Experiments in Speech Emotion Recognition

Performance of the speech emotion recognition systems depends on many factors such as quality of the speech data, environment, cultural differences, language, emotion categorization scheme, etc. In this work, we create a baseline speech emotion recognition model based on convolutional neural networks using the RAVDESS dataset. First, we compare the performance of the model with different labeling schemes. Then, we perform cross-corpus experiments on datasets recorded in different languages. The results show that emotion groups with common arousal or valence categories are often confused and using multiple corpora in training improves the generalization capacity of the model.