Using a Combined Strategy to Improve the Accuracy in Predicting the Pronunciation of the Polyphonic Character 'chang' in Chinese Texts

Autor: Chien-Ting Li, 李建廷
Rok vydání: 2019
Druh dokumentu: 學位論文 ; thesis
Popis: 107
In brief, in the processing of Text-To-Speech, the first step is to get enough information from the text to determine the pronunciation. The second step is to obtain the relevant prosodic information. Finally, we use possible ways to generate waveforms. For the first step mentioned above, in Chinese, the current results are still not bad; but the problem of polyphonic characters still needs to be improved. Therefore, this thesis conducts related research on the sentence after the word segmentation. We focused on the polyphonic character word "chang". We use a single feature combination using 9 single features and confidence. At the same time, a total of 36 groups were also explored by using paired features. We expect to get enough information in a limited corpus. We extended single features to paired features, hence there are more features that we can use. With these features, we designed some steps to determine the pronunciation. We measured the accuracy of each step and put the steps with higher accuracy in the front. One of the steps is that after the score of each feature was sorted, we looked the pronunciation of some features with the highest scores. The pronunciation of “ㄔㄤˊ” will be selected only if it has more than one supporting feature. This is the way we adopted in response to the fact that the ratio of the pronunciation of “ㄔㄤˊ” is higher than that of “ㄓㄤˇ”. The accuracy of the outside test is 95.42%.
Databáze: Networked Digital Library of Theses & Dissertations