Synthetic Speech Evaluation by 2D GMM Classification in Pleasure-Arousal Scale

The paper is focused on a description of a system for automatic evaluation of synthetic speech quality based on two-dimensional detection in the Pleasure-Arousal (P-A) scale. The original speech material of a speaker used for synthesis is compared with the synthesized one to find similarities/differences between them. For continual P-A detection, the Gaussian mixture model (GMM) classifier is used. The GMM models of the P-A classes are created and trained using the sound/speech material from the database labelled directly in the P-A scale without any relation with the used original speech or the tested sentences. The basic experiments confirm the principal functionality of the developed system. Additional analysis shows the great importance of the proper selection of the number of mixtures, and the used type of the sound/speech database for GMM models building. The obtained objective evaluation results are highly correlated with the subjective ratings of human evaluators.