An Initial Study on Hakka Speech Synthesis

• Main Authors: Shie-Jen Lee, 李雪貞
• Other Authors: Hung-yan Gu
• Format: Others
• Language: zh-TW
• Published: 2002
• Online Access: http://ndltd.ncl.edu.tw/handle/45818970175515086055

碩士 === 國立臺灣科技大學 === 資訊工程系 === 90 === In this thesis, speech synthesis for the sub-dialect of Hakka, Hai-Lu, is studied. Since Hai-Lu Hakka has 7 lexical tones and 736 base syllables, it is more complex than other Hakka sub-dialects. Signal waveform synthesis and prosody parameter generation are the focus of this thesis. For signal waveform synthesis, the method, Time-Proportioned Interpolation of Pitch waveform (TIPW), is adopted and modified to solve the synthesis problem that some Hakka syllables are pronounced with abrupt-tone. Abrupt-tone is not found in Mandarin. For prosody parameter generation, we suppose that the sentence pitch-contour hidden Markov model trained for Mandarin speech synthesis could be directly borrowed to generate pitch-contour parameters for a Hakka sentence if an appropriate tone mapping rule is used. Therefore, we had studied this problem and proposed a Hakka-to-Mandarin mapping rule. According to the perception experiments conducted, the evaluation results show that our supposition is workable. About the parameters of syllable duration and amplitude, generation rules proposed by previous studies are adopted and modified here to have better prosody expression. Using the strategies mentioned above, we had built a prototype system for Hakka speech synthesis. Some simple processing functions for text-analysis are also implemented. With this system, two sets of sentences are synthesized for evaluation experiments. One set of Hakka sentences are synthesized earlier than the other set of Min-Nan sentences when the generation rules for prosodic parameters are not settled. Due to this timing difference and using of a notebook computer’s poor speaker for playing, the comprehension and naturalness scores of Hakka synthetic speech only reach 91.87% and 79.5, respectively. But for the set of Min-Nan synthetic speech, the scores obtained are 97.1% and 85.5 for comprehension and naturalness, respectively.