基於決策樹演算法之台語連音變調預估模組

A Prediction Module for Taiwanese Tone Sandhi Based on the Decision Tree Algorithm

台語連音變調問題為研究台語文轉音系統的重要問題之一。在詞的階層，大多數的詞都遵循詞尾本調，非詞尾變調的一般變調規則。在句子的階層，台語的連音變調問題會變得比較複雜，因為一般變調規則並不能完全適用在句中的每個詞上。在這篇論文中，我們提出了一套可用於中文文句轉台語語音系統的台語連音變調預估模組。我們以決策樹C5.0演算法搭配三種Speaal Case來對句子中的各個音節做連音變調預估，此預估模組在內部測試和外部測試的預估正確率分別為93.42%和91.13%。

Taiwanese tone sandhi problem is one of the important research issues for Taiwanese Text-to-Speech systems. In word level, we can use the general tone sandhi rules to deal with the Taiwanese tone sandhi problem. The tone sandhi becomes more difficult in sentence level because of that the general tone sandhi rules for words may not apply at each word in a sentence. In this paper we proposed a module to deal with the Taiwanese tone sandhi problem for Chinese to Taiwanese Text-to-Speech systems. We adopt Decision tree C5.0 algorithm accompanied with three Special Cases generated from training data to predict the tone sandhi of each syllable. In this module, the accuracy of the inside test and outside test are 93.42% and 91.13%, respectively.