Synthetic Speech Evaluation by Differential Maps in Pleasure-Arousal Space

摘要

The paper deals with automatic evaluation of the quality of synthetic speech using Gaussian mixture models (GMM) for classification in the Pleasure-Arousal (P-A) scale and subsequently calculated 2D and 3D P-A differentials maps. The speech synthesized from the voice of a speaker is compared with the original voice of the same speaker. Three methods of speech synthesis are ordered by descending 3D perceptual distances from the original speech material. Basic experiments confirm the principal functionality of the developed system. The detailed analysis shows a great influence of the number of mixture components, the size of the processed speech material, and the type of the database for GMM creation on partial results of the continual P-A detection and the final results. The objective evaluation results are finally compared with the subjective ratings by human evaluators.