论文部分内容阅读
近年来,中文文语转换系统趋向于采用基于语料库上下文的单元选择法来获取拼接的语音单元实例,事实证明这种方法对于合成高质量的语音非常有效。单元选择的关键是各单元特征权值的确定,这里将提出一种非常有效的在单元选择中确定各特征权值的方法。它是对现有的利用线性回归确定权值方法的改进。其主要思路是对单元进行分类,对于各类单元分别进行线性回归,同时定义了一个客观衡量语音差异的距离,从而可以回归得到各特征的恰当权值。
In recent years, the Chinese text-to-speech conversion system tends to adopt the unit selection method based on the corpus context to obtain spliced speech unit instances, which has proved to be very effective in synthesizing high-quality speech. The key to cell selection is to determine the weight of each cell feature. Here we present a very effective way to determine the weight of each cell in cell selection. It is an improvement over the existing method of determining weights using linear regression. The main idea is to classify the cells, and to linearly regress each type of cells separately, meanwhile define a distance that objectively measures the difference of speech, so that the appropriate weight of each feature can be regressed.