基于概率频度的普通话韵律结构预测统计模型

Statistical model based on probability frequency for Mandarin prosodic structure prediction

为进一步提高文语转换系统中韵律结构预测的准确度,提出了一个基于概率频度的统计模型的方法,预测韵律词和韵律短语边界两级韵律结构。该方法提取与韵律词和韵律短语边界有关的语言学特征(词性、语法词、长度和位置等),并进行样本训练计算各个特征的概率频度值,最终分别建立韵律词和韵律短语的统计模型。实验结果表明:统计模型的方法对于韵律词和韵律短语边界预测的正确率分别可达90.6%和84.6%,并与决策树算法和T ransform ation-based learn ing(TBL)转换规则学习算法比较,提高10%以上的正确率。

The accuracy nf prosody structure prediction in text-to-speech （TTS） conversion systems is improved by a statistical model based on the probability frequency to detect the two-tier prosodic hierarchy, including prosodic words and prosodic phrases. The system fast extracts linguistic features related to the prosodic structure such as part of speech, lexical words, length, and position information, Then, the probability frequency for each selected feature is calculated with statistical models designed for the prosodic words and phrases. Tests show that the correct identification rates of prosodic words and phrases are improved to 90.6% and 84.6% using the statistical model. The statistical model gives 10% better performance than the decision tree Transformation based learning （TBL） algorithms.