Summary: | 碩士 === 國立中興大學 === 資訊網路多媒體研究所 === 99 === This thesis aims at the implementation for Hakka Text-to-Speech (TTS) System on Internet. Our system is composed of four components as follows: Text analysis, Mandarin to Hakka, Prosody prediction, and Speech generation module. More than 5400 monosyllabic speech units and 4063 word speech units of Hakka and several silences with various durations have been recorded as basic unit for speech synthesis. By adding breaks to Hakka sentences and finding out the pronunciation of polyphonic characters appropriately, we can provide real synthesis speech with frequent, prosodic and natural quality on Internet .
We focus on solving pronunciation ambiguity of polyphonic characters, i.e., to determine which pronunciation should be chosen. We predict pronunciation by using Bayesian network classifier、 C4.5 decision tree classifier、 CART classifier, and SVM classifier. The result of our experiments show that we can handle the prediction of some words very well in our Hakka Text-to-Speech System.
|