Research on Synthesis Method for Spoken Sentences from Knowledge Representation

本项目以构建一个将机器可读的知识表示转化为连续语音的语音输出系统为目标，研究内容包括概念表示、语言表示、音韵符号、语音波形的生成以及整个系统的构建。该项目几乎完全完成，并获得了以下结果：本课题旨在实现一个从知识库中为用户选择所需信息的问答系统。在系统中引入对话控制过程，可以有效地为用户精确提取所需信息，实现系统与用户之间更流畅的界面，生成椭圆/代词表示，控制焦点条件。通过上述过程提取的知识表示，提出了一种构建深层语义表示的方法，深层语义表示基本是用例结构。并在计算机上编写并实现了一个程序，利用该语义表示生成表层句子。提出并在计算机上实现了用句子边界、连接子句边界、子句边界和ICRLB（递归左分支结构的直接成分）边界表示句法结构，用韵律词重要性三级指标表示语篇结构的方法。该方法进一步利用韵律词的句法和语篇信息，以及韵律词的语法性质和重音类型，生成电话符号和韵律符号。在计算机上研制并构建了一种新型的终端模拟语音合成器，该合成器由四条独立的级联路径组成，分别模拟四种不同语音类别和声门波形、摩擦噪声和爆炸噪声三种激励源的语音传递函数。基于自然语音分析的结果，为上述语音合成器构建了所有可能的日语音节模板。构建了一个程序，将存储的音节模板连接起来，生成连续语音特征参数的时间模式。将上述结果进行总结，形成了一个完整的语音输出系统，并通过合成语音的听力测试验证了其有效性。少

With the aim of constructing a speech output system which converts machine readable knowledge representation into continuous speech, this projects includes the studies on the generation of concept representation, linguistic representation, phone and prosodic symbols, speech waveform, and the construction of the total system. The project was almost fully performed and following results were obtained.1. The study was performed to realize a question answer system which selects necessary information for the users from the knowledge base.2. The introduction of the dialogue control process to the system was shown to be effective for precisely extracting necessary information for the users, for realizing smoother interface between the system and the user, for generating elliptical/pronominal representation, and for controlling focal condition.3. A method was developed to construct deep level semantic representation, which is basically case structure, from the knowledge representation extracte … More d by the above procedure. A program was also constructed and realized on the computer which generates surface level sentences from this semantic representation.4. A method was developed and realized on the computer which represents syntactic structure using terms of sentence boundary, concatenated clause boundary, clause boundary, and ICRLB (immediate constituent with recursively left-branching structure) boundary, and discourse structure using three-level indices of importance of the prosodic words. This method further generates phone and prosodic symbols from the syntactic and discourse information, and grammatical property and accent type of the prosodic word.5. A new type of terminal analog speech synthesizer was developed and constructed on the computer which consists of four separate cascade paths simulating the vocal transfer functions of the four different speech categories and three types of excitation sources, viz., glottal waveform, fricative noise and plosive noise.6. Syllable templates for above mentioned speech synthesizer were constructed for all possible Japanese syllables based on the results of analysis of natural speech. A program was also constructed which concatenates stored syllable templates and produces temporal pattern of feature parameters of continuous speech.7. The above results were summarized to produce a total speech output system and its validity was proved by the listening test of synthesized speech. Less