The Prediction of Pronunciation of Polyphonic Words in a Chinese to Taiwanese TTS System

system. There are two kinds of polyphonic problem in a Chinese to Taiwanese TTS system: polysemy problems and separated words. Polysemy refers to words with more than one meaning or pronunciation. For example, there are two kinds of pronunciation for the word “他” (he) in Taiwanese. They are /yi7/ and /yin7/. The first pronunciation, /yi7/, can mean ‘he’ or ‘him’; and the other one, /yin7/, means ‘his’. Separated words refer to situations when a fixed phrase or expression is inserted with descriptive wordings because of a particular writing habit. To illustrate, the sentence “洗一個舒服的澡” (taking a comfortable bath), for instance, will be read as “洗 /se2/ 一個爽快的澡/chhau2/” in Taiwanese, rather than the accurate one, “洗 /se2/ 一個爽快的身軀 /sin7-khu7/』, if the system is not able to recognize that “洗”(wash) and ”澡”(bath) are actually taken from a fixed phrase. The correct pronunciation of a word affects the comprehensibility (or clarity) and fluency of Taiwanese speech. We propose a combined approach to polysemy problem and an aproach of using the semantic roles revealed in sentence parsing trees to judge whether or not words belong to a certain fixed phrase. Results show that the two approaches can solve the polyphonic problems efficiently.