Improving prosodic phrasing of Vietnamese text-to-speech systems

End-to-end TTS architecture which is based on Tacotron2 is the state-of-art system. It breaks the traditional system framework to directly converts text input to speech output. Although it is shown that Tacotron2 is superior to traditional piping systems in terms of speech naturalness, it still has many defects in building Vietnamese TTS: 1) Not good at prosodic phrasing for long sentences, 2) Not good at expression for foreign words. In this paper, we used 2 methods to solve these defects: 1) Pause detection system for predicting and inserting punctuation into long sentences to improve speech naturalness. 2) Translation system for transcribing foreign words to Vietnamese words. In the VLSP 2020 evaluation campaign, our model achieved a mean opinion score (MOS) of 3.31/5 compared to 4.22/5 of humans.