Research on a speech synthesis system for Japanese texts based on the natural language understanding

本课题的目标是构建一个日语正字符号文本到标准日语发音的连通语音转换系统。取得的主要结果如下：(1)提出了一种对天气预报句子进行句法和语义分析并将结果转换为适合语音合成的形式的方法。(2)利用基频轮廓生成模型定量分析了广播的基频轮廓特征，特别是句子的句法结构与基频轮廓短语成分之间的关系。(3)在分析连通语音中重音成分的基础上，定量地阐明了重音成分与词的重音类型、句子的句法结构和语篇结构之间的关系更多的构造规则用于生成没有在文本中明确表示的韵律符号。这些规则利用标点符号、单词的重音类型、句子的句法结构和篇章的语篇结构来生成符号。(5)基于听力测试的结果，改进了存储的音节模板的质量。(6)研究了存储的音节模板的ARMA参数的拼接方法。在上述研究成果的基础上，结合上一年的研究成果，开发了一套日语文本语音合成系统。在该系统中，分段特征通过拼接存储的音节模板来合成，而韵律特征则利用基频轮廓生成模型按规则合成。通过对日语文本合成语音的听力测试，验证了该系统的有效性。虽然该系统最初是针对天气预报句子的，但它也适用于其他领域的句子。以上结果表明，该项目已基本完成。较少

The aim of this project is to construct a system which converts a Japanese text of orthographic symbols into connected speech with pronunciation of the standard Japanese.The main results obtained are as follows:(1) A method was developed which conducts syntactic and semantic analysis of weatherforecast sentences and transforms the results into a form appropriate for the speech synthesis.(2) Characteristics of fundamental frequency contours of announcements were analyzed quantitatively by using the model for fundamental frequency contour generation with special emphasis on the relationship between the syntactic structure of a sentence and the phrase component of a fundamental frequency contour.(3) Based on the analysis of the accent component in connected speech, relationships were quantitatively clarified between the accent components and the accent type of a word, the syntactic structure of a sentence and the discourse structure of a text.(4) The results in (1) to (3) were combined to … More construct rules for generation of prosodic symbols which are not explicitly represented in the text. These rules generate the symbols using the punctuation marks, the accent type of a word, the syntactic structure of a sentence and the discourse structure of a text.(5) The quality of the stored templates of syllables was improved based on the results of a listening test.(6) Investigations were made on the method for concatenating ARMA parameters of the stored syllable templates. The subjective evaluation was conducted for the synthesized speech.(7) Based on the above-mentioned results and those obtained in the preceding year, a system for speech synthesis of Japanese texts was developed. In this system, the segmental features are synthesized by concatenating stored syllable templates, while the prosodic features are synthesized by rule using the model for fundamental frequency contour generation. The validity of the system was proved by the listening tests of synthesized speech of Japanese texts. While this system was originally aimed at the weatherforecast sentences, it was also valid for sentences of other fields.The above results indicate that the project was almost fully accomplished. Less