Design And Implementation Of Indonesian Speech Synthesis System Based On HMM

Bahasa Indonesian(referred to as Indonesian)originates from Malay in the northeastern part of Sumatra.It belongs to the Malay-Polynesian language.More than 30 million people worldwide speak Indonesian as their mother tongue,and about 100 million people speak Indonesian as a second language.Modern Malay and Indonesian are spelled in Latin,and the spelling of the two languages is similar.In order to develop the Indonesian language translation application system,this paper designs and implements the Hidden Markov Model(Hidden Markov Model,HMM)-based Indonesian language conversion baseline system,and on this basis,explores ways to improve the naturalness of speech synthesis.The main work of the thesis includes:(1)Automatic segmentation of phonemes.According to the phonetic characteristics of Indonesian language,two kinds of phonetic subsets are determined according to the structure of the Initials and finals and the phoneme structure.And using HMM-based automatic segmentation technology,based on two different sets of phonemes,the automatic segmentation of Indonesian speech is carried out,which lays a foundation for further research on the subsequent texts.(2)Implementation of model training and speech synthesis.Based on HTS(HMM-based speech synthesis system),a complete speech synthesis system framework is established.Firstly,tri-phones' context attribute and problem set are designed,and the acoustic model is trained by decision tree clustering,and finally the Indonesian speech synthesis is realized.(3)Improvement of the Indonesian speech synthesis system.This paper has done three improvements to the Indonesian speech synthesis system.Firstly,for the phenomenon of zero initials in Indonesian,this paper designs a zero-acoustic proton that conforms to the acoustic performance of Indonesian and matches the training system,and implements the synthesis system after introducing zero initials.Secondly,becauseIndonesian is a language written in Latin,the phoneme is a basic part of Indonesian.Therefore,this paper considers Indonesian phoneme as a synthetic primitive,and designs a phonetic synthesis system based on phonemes and tri-phones.The correct rate of the segmentation of the sound is improved,and the quality of the speech synthesis is improved.Finally,this paper designs a synthesis system based on phoneme and full-context lab.The context attributes and problem sets are designed,and the acoustic model training is carried out,and finally a high quality speech is synthesized.The experimental results show that when the phoneme is selected as the synthetic primitive for automatic segmentation of the sound,the correct rate of segmentation reaches 89.36%,which is improved by 13.04 compared with the previous note selected according to the structure of the Initials and finals.For the speech synthesis results,the improved system's synthesized speech has also been greatly improved in terms of naturalness and accuracy..