JISE


  [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13]


Journal of Information Science and Engineering, Vol. 40 No. 2, pp. 283-302


Development of a Taiwanese Speech Synthesis System Using Hidden Markov Models and a Robust Tonal Phoneme Corpus

YUNG-JI SHER1,+, MING-CHUN HSU2, YU-HSIEN CHIU3, YEOU-JIUNN CHEN5,
CHUNG-HSIEN WU4 AND JIUNN-LIANG WU6
1Department of Special Education and Graduate Institute of Rehabilitation Counseling
National Taiwan Normal University
Taipei, 106 Taiwan
2Wistron Neweb Corporation
Hsinchu, 300 Taiwan
3NSTC AI Biomedical Research Center
4Department of Computer Science and Information Engineering
National Cheng Kung University
Tainan, 701 Taiwan
5Department of Electrical Engineering
Southern Taiwan University of Science and Technology
Tainan, 710 Taiwan
6Department of Otorhinolaryngology, National Cheng Kung University Hospital
College of Medicine, National Cheng Kung University
Tainan, 704 Taiwan
E-mail: siaa@ntnu.edu.tw1,+; smjeremy@gmail.com2; chiuyuhsien@gmail.com3;
chenyj@stust.edu.tw5; chwu@csie.ncku.edu.tw4; jiunn@mail.ncku.edu.tw6

The number of young native speakers of Taiwanese, the variant of Southern Min spoken in Taiwan, has decreased. Technological advancements such as text-to-speech (TTS) systems could help arrest this decline. The aim of this study was to design a robust tonal phoneme corpus and a speech synthesis system for Modern Literal Taiwanese (MLT). MLT subsyllables were analyzed using phonetics and phonology to establish tonal phoneme models. These robust tonal phoneme models and hidden Markov models were used to construct an MLT TTS synthesis system. Algorithm-based training resulted in 869 balanced sentences containing 12,544 syllables, with each sentence containing an average of 14.4 syllables. In total, 218 sentences, which included rare phonemes, were manually drafted to supplement the corpus. The synthesized phonemes were deemed to have high intelligibility and could be included in the developed TTS system. According to the HTK speech recognition tool, the overall phoneme recognition rate was 96.47%. Testers, who were native Taiwanese speakers, assigned the synthesized sentences a mean opinion score of 4, indicating that they sounded natural. This developed system and the results described herein can inspire future developments in speech technology and computational linguistics.

Keywords: Taiwanese, text-to-speech, speech corpus, hidden Markov model, modern literal Taiwanese

  Retrieve PDF document (JISE_202402_06.pdf)