Phoneme segmentation using deep learning for speech synthesis

Abstract

In this paper, we propose the phoneme segmentation method, which is one of the basic module that consist unit-selection-based speech synthesis, using deep learning algorithm. To enhance this, we apply the additional cross entropy loss into the Deep speech based speech recognition architecture. From this approach, we can get higher accuracy of phoneme boundary. In our experiments, the proposed method has 20.91 % boundary accuracy which is higher than the conventional phoneme segmentation.