英语语音合成中基于有限泛化法的字素切分规则的机器学习

English grapheme segmentation rules learning based on the finite generalization algorithm 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,因此不可能创建包含所有词汇的词库。对于未包含在词库中的英语单词,通过“字母转换成音素(L2P)”算法自动生成其音标是一个最好的解决办法。而L2P首要的任务就是字素切分。为此,文中提出了一种有限泛化法(FGA)的机器学习算法,用于进行字素切分规则学习。用于学习的词典库有27 040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,学习实例和测试实例的平均实例切分正确率为99.84%和97.88%,平均单词切分正确率为99.72%和96.35%;平均规则数为472个。

Letter-to-Phoneme Conversion（L2P） is a very important component in English speech synthesis system. The first task of L2P is grapheme segmentation. A machine learning method named the Finite Generalization Algorithm （FGA） was presented, which was used to learn rules of English grapheme segmentation. The average accuracies of training and testing sets were 99.84% and 97.88% respectively for instances segmentation, and 99.72% and 96.35% respectively for words segmentation. The average number of rules is 472, about 1 rule per 52 words.