Shaanxi Normal University, Ph.D (2016)ĭu, G.: Research on Tibetan speech synthesis technology based on neural network. Northwest University for Nationalities, M.S (2014)Ĭairang, Z.: Research on Tibetan speech synthesis technology based on mixed primitives.
Zhang, J.: Research on Tibetan Lhasa speech synthesis based on hmm. Sun, Y., et al.: Tibetan question generation based on sequence to sequence model. The results of experiments show that the effect of synthesized speech is improved with dialect ID. In the end, the dialect ID is appended to the text to help the model learn the characteristics of three dialects during training. Then an end-to-end multi-dialect shared speech synthesis model is established for Amdo pastoral dialect, Dege-Kham dialect and Yushu-Kham dialect, which maps the Latin transliteration vector of Tibetan character to Mel spectrograms. First the Wiley transliteration scheme is used to convert Tibetan characters into corresponding Latin letters, which reduces the size of the training corpus, reduces the workload of front-end text processing and improves the efficiency of speech synthesis. This paper proposed an end-to-end multi-dialect Tibetan speech synthesis technique based on the Tacotron2 model.
The end-to-end speech synthesis model, Tacotron2, which is based on neural network theory, drives advancements in speech synthesis technology.
0 kommentar(er)
