英语语音合成中基于DFGA的字音转换算法

Learning English Grapheme-to-Phoneme Conversion Rules Based on DFGA in English Speech Synthesis

在英语语音合成中,由于英语有着几乎无限多的词汇,我们不可能创建包含所有词汇的词库。因而对于未包含在词库中的英语单词,通过“字素转换成音素(G2P)”算法自动生成其音标是一个最好的解决办法。为此,论文提出了一种动态有限泛化法(DFGA)的机器学习算法,用于进行字素/音素转换规则的学习。用于学习的词典库有27040个单词,其中90%的词用于规则学习,剩下的10%用于测试。经过10轮交叉验证,学习集和测试集的平均字素转换正确率为99.78%和93.14%,平均单词转换正确率为99.56%和73.51%。

Grapheme-to-Phoneme Conversion（G2P） is a very important component in English speech synthesis system. This paper presents a machine learning method named the Dynamic Finite Generalization Algorithm（DFGA）,which is used to learn rules of English grapheme-to-phoneme conversion.The average accuracies of training and testing sets are 99.78% and 93.14% respectively for graphemes conversion,and 99.56% and 73.51% respectively for words conversion.