Artefact Determination by GMM-Based Continuous Detection of Emotional Changes in Synthetic Speech

The paper is focused on a description of a system for automatic detection of speech artefacts based on the Gaussian mixture model (GMM) classifier. The system enables to detect one or more artefacts in synthetic speech produced by a text-to-speech system. Our speech artefact detection uses continual GMM classification of emotional states in 2-D affective space of valence and arousal within the whole sentence and calculates the final change in the evaluated emotions. The detected shift to negative emotions indicates presence of an artefact in the analysed sentence. The basic experiments confirm functionality of the developed system producing results with sufficient correctness of artefact detection. These results are comparable to those attained by a standard listening test method. Additional investigations show relatively great influence of the number of mixtures, the number of used emotional classes, and types of speech features on the evaluated emotional shift.