Nowadays a corpus-based unit concatenation text-to-speech (TTS) system has been widely used because of its high quality synthesized speech. The high quality synthesized speech in a corpus-based TTS is obtained by using a large amount of speech DB in i...

Nowadays a corpus-based unit concatenation text-to-speech (TTS) system has been widely used because of its high quality synthesized speech. The high quality synthesized speech in a corpus-based TTS is obtained by using a large amount of speech DB in implementing the system. However, it is a difficult job and costs very much to collect a phonetically balanced large amount of speech DB, and segment to extract synthetic units having various voice characteristics. Thus, it is generally used for the serve based TTS system and is hard to be applied to the embedded system such as mobile devices having the limitation of the memory size. On the other hand, an HMM-based text-to-speech system (HTS) has recently drawn much attention to overcome such a problem. The HTS uses the statistical model, hidden Markov model (HMM) as a synthetic unit, to represent the spectra and prosodic characteristics of the speech signal. Thus the synthesis engine needs less memory and low computation complexity and is suitable for the embedded system. It also has the advantage that voice characteristics of the synthetic speech can be modified easily by transforming HMM parameters appropriately.
In this thesis, we implemented an HMM-based Korean text-to-speech system using a small sized Korean speech DB. We used the HTS software released on the Internet website with some amount of ETRI 611 DB and SeoulMal DB. The ETRI 611 DB, phoneme labeled 611 words originally made for training the speech recognition system, was used to generate initial HMMs that represent context-independent monophone acoustic models. With the monophone HMMs, then, SeoulMal DB was used to generate context-dependent triphone HMMs. We used the {preceding, current, succeeding} phonemes, position of the current phoneme in the current phrase and the number of syllables in the current phrase as contextual factors to model context-dependent HMMs.
The synthesized speech has shown very intelligible vocoded speech quality though naturalness was not enough. This is because, we think, prosodic feature parameters were not modeled well in the HMM training procedure due to the limited speech DB. Thus we increased naturalness of the synthesized speech a little by simply controlling the pitch pattern of the phrase and sentence. The file size of the implemented HMM-based Korean text-to speech system was about 1.3 Mbytes, so it could be used for the embedded system.

•   영어권 학습자를 위한 한국어 교재 구성

•   깔뱅의 기도론 연구

•   高职院校韩语系建设的几点思考

•  韩国跆拳道运动的文化价值观探讨

•   항공사의 지각된 서비스품질이 실용적

•   영어 문장구조에 대한 이해가 읽기와 듣

•   형태 초점 접근법을 활용한 한국어 대조

•   한·중 사동 표현의 대조 연구

•   모야모야 환아의 수술 후 자기효능감,

•   중국인 학습자를 위한 한국어 거절 화행

•   도시지역 여성결혼이민자의 재사회화

•   한국과 독일의 중등교육단계에서의 진로

•  汉韩常用颜色词对比探讨

•   TV 포맷의 새로운 유형화 : 이야기, 놀이

•  韩国电影剧本中会话含义的略论探讨