A study on acoustic parameter selection strategies to improve deep learning-based speech synthesis

In this paper, we investigate the variation in the performance of a deep learning-based speech synthesis (DLSS) system based on the configuration of output acoustic parameters. Our method is mainly applicable for vocoding-based statistical parametric speech synthesis (SPSS), which has advantages in low-resource scenarios. Given the independence assumption of the source-filter model for the spectral and fundamental frequency F0 parameters, we propose a reliable network architecture for training acoustic parameters. Particularly, the F0 parameter suffers from high fluctuation and an extremely low number of dimensions. To relieve these problems, we introduce a context-window approach. Furthermore, we apply data augmentation to the proposed structure to overcome a lack of training data, which is a frequent issue with multi-speaker TTS systems. Experimental results confirm the superiority of the proposed algorithm over conventional ones in both single-speaker and multi-speaker TTS setups.