語音合成是一個急劇增長的需求領域，舉凡我們每天使用的手機、電腦、智慧手錶甚至是自駕車都需要這個技術。然而，雖然世界上有很多人在各種語言上進行語音合成的開發，身為臺灣母語之一的台語卻鮮少有人為其發展相關技術。
因此，為了解決台語語音合成的技術痛點，我們從三個方向著手進行嘗試，分別為中文到台語翻譯、台語變調處裡以及台語語音合成。其中各自遭遇到的主要問題分別為一字多音、轉調規則覆蓋與系統成型，但本論文提出各種方法解決，並聘請台語語言學專家協助提供台語的變調規則以及台語語音合成方面提供決策樹檔案的協助，最終開發出了一個基於變調的台語語音合成系統與中台翻譯應用系統，其包含了詞性與頻率選音的中文到台語翻譯模組、決策樹變調模組與NTPU和Kaldi-HTS兩個台語語音合成模組，能夠進行基礎的中文對台語翻譯。
實驗結果我們在中文轉台語翻譯方面得到了91%的正確率、轉調方面得到了91.8%的正確率、並製作了兩個可用的台語語音合成系統，而我們實作出來的兩個合成系統能夠發出更加道地的台語語音。

Speech synthesis is a rapidly growing demand area. For example, the mobile phones, computers, smart watches and even self-driving cars we use every day need this technology. However, although there are many researchers in the world have developed speech synthesis technology in various languages, however, Taiwanese, one of Taiwan’s mother tongues, few researchers have developed speech synthesis for it.
Therefore, in order to solve such technical pain points of Taiwanese speech synthesis, including three issues, Chinese to Taiwanese translation, Taiwanese tone sandhi, and Taiwanese speech synthesis. The main problems encountered by each of them are polyphony, the conflict of tone sandhi rules, and how to construct the system of Taiwanese speech synthesis, but these problems are all solved in this thesis. First, we ask Taiwanese linguistic experts to assist in providing Taiwanese tone sandhi rules and providing the decision tree file of Taiwanese speech synthesis. Finally, we developed a Taiwanese Speech Synthesis System based on Taiwanese Tone Sandhi and implemented a translation system of Chinese to Taiwanese, including Chinese to Taiwanese translation module based on frequency and part-of-speech phonetic selection, decision tree based tone sandhi module, NTPU and Kaldi-HTS Taiwanese speech synthesis modules. Combining the three modules can assist users obtain translations from Chinese to Taiwanese and generate fluent Taiwanese speech synthesis.
We conducted three experiments and obtained effective results. First, Chinese to Taiwanese translation, we achieved a correct rate of 91%, and for Taiwanese tone sandhi module, we achieved a 91.8% correct rate, and both Taiwanese speech synthesis systems are feasible, and can generate more authentic Taiwanese voice.

[1]. http://www.twlls.org.tw/jtll/documents/13.2-4.pdf 臺灣歷次語言普查回顧, 台灣語文學會
[2]. 蔡宗謀. 中文文句轉台語語音系統初步研究. 中興大學資訊科學與工程學系所學位論文, 2008, 1-47.
[3]. WANG, Jhing-Fa; WHANG, Bau-Jang; LIN, Shun-Chieh. 國語文句翻台語語音系統之研究 (A Study for Mandarin Text to Taiwanese speech System)[In Chinese]. In: Proceedings of Research on Computational Linguistics Conference XII. 1999. p. 37-53.
[4]. 許文漢, et al. 基於深度學習之中文文字轉台語語音合成系統初步探討. 中文計算語言學期刊, 2020, 25.2: 69-83.
[5]. 李尚德. 台語辭典建構與台語變調探討. 中興大學資訊科學與工程學系所學位論文, 2007, 1-85.
[6]. 許書豪. 台語連音變調問題研究. 中興大學資訊網路多媒體研究所學位論文, 2010, 1-51.
[7]. 楊允言, et al. 台語變調系統實作研究. 第十七屆自然語言與語音處理研討會論文集, 台南, 台灣, 2005, 293-304.
[8]. 蔡珮均. 基於決策樹演算法之台語連音變調預估模組. 中興大學資訊科學與工程學系所學位論文, 2012, 1-54.
[9]. 楊鈺清, et al. 台語文句翻語音系統之製作. 1998. PhD Thesis.
[10]. 黃競億, et al. 台語 TTS 變調規則與斷詞器之製作. 2000. PhD Thesis.
[11]. 鐘祥睿, et al. 台語 TTS 系統之改進. 2001. PhD Thesis.
[12]. https://github.com/ChhoeTaigi/ChhoeTaigiDatabase ChhoeTaigi 找台語：台語字詞資料庫
[13]. https://ithuan.tw/ 意傳科技官網