摘要： 在基于语料库的文语转换系统中，需要事先选择并合成一定数量的句子得到其MOS评分，通过使拼接损失与MOS分相关系数最大来优化拼接损失函数。该文以清华大学TH-Coss女声语料库为例，提出了归一化语境误差和与语料库相似度的概念，以归一化语境误差和最小为准则，从TH-Coss语料库中选出应该事先合成的那批句子，用剩下的句子替代整个语料库。实验结果表明，与随机选取相比，该文的方法选取出来的句子同整个语料库的相似度有较大幅度的提高，剩余句子组成的语料库与整个语料库的相似度达到0.995 5，表明这种替代是合理的。

关键词: 相似语料库, 归一化语境误差和, 语料库相似度, 文语转换系统, 拼接损失函数

Abstract: To optimize the cost function for unit selection in the corpus-based TTS system, a certain number of sentences should be preselected and synthesised. These sentences’ mean opinion scores (MOS) can be got so as to maximize the corrrelation between the concatenative cost and the MOS. This paper proposes the concepts of sum of normalized contextual errors (SNCE) and corpus-similarity (CS) based on the TH-Coss corpus of women provided by Tsinghua University. It gets the preselected-sentences and uses the rest of the sentences to replace the TH-Coss corpus using the guideline of minimizing SNCE. Experimental results show that the CS between the preselected-sentences and the TH-Coss corpus rises greatly. And it is valid to replace the TH-Coss corpus with the rest of the sentences for the CS between them reaches 0.995 5.

Key words: similar corpus, sum of normalized contextual errors(SNCE), corpus-similarity(CS), text-to-speech (TTS), concatenative cost function

中图分类号: 

• TP393