Rule-Synthesis of Spoken Sentences for the Speech Dialogue Systems

组织这一项目的目的是建立一种技术，通过规则合成具有高自然度的高质量语音，以回答以适合计算机处理的知识表示(深层语义表示)的形式提供的问题。韵律的自然性可以通过考虑整个话语的语言信息来获得。具体地说，构建了一个由对话处理和语音合成部分组成的语音响应系统，作为语音对话系统输出生成的主要组成部分。以下结果表明，该项目的原定目标已经达到。在选择滑雪场指南作为语音应答系统的主题后，利用商业可用的数据库构建了系统的必要要素词典和知识库。对话处理规则分为两组，一般规则为…根据主题和任务，提供更多对话流程和具体规则。每个主题相关规则都是基于每个对话样本构建的。根据上述结果，开发了一个对话处理系统，并对其运行情况进行了检查。结果表明，通过引入话题相关规则，只有输入语音中相关词的信息才能生成应答，而输入语音中的省略和回指可以很容易地通过参考先前使用的话题相关规则来补偿。主题相关规则的优点还表现在易于添加和修改规则。开发了一种从对话处理获得的响应内容生成表层句子的方法。反应内容以格结构表示的列表形式表示，表层句子通过加助词和谓词共轭生成。在这个句子生成的过程中，音素和韵律符号也是利用语言信息生成的。对拟对话语音的韵律特征进行了分析。基于结果，修改了先前为朗读语音开发的韵律规则，以产生用于对话语音的韵律规则。这些规则包括焦点控件。将上述结果与已开发的语音合成系统相结合，构建了一个语音响应合成系统。进行了听力测试，证明了系统的有效性。为了提高语音质量，还研制了语音合成器。较少

This project was organized with the aim of establishing a technology for synthesizing by rule a high-quality speech with high-naturalness in prosody as a response to the question supplied in the form of knowledge representation (deep-level semantic representation) suitable for the computer processing. The naturalness in the prosody may be obtained by taking the linguistic information of the whole utterance into account. To be concrete, a speech response system of dialogue processing and speech synthesis parts was constructed as a major component for the output generation in the speech dialogue systems. Following results indicate the original goal of the project being attained.1. After selecting the guidance of the ski areas as the theme of the speech response system, a word dictionary and a knowledge base, necessary elements of the system, were constructed utilizing the commercially avail- able database.2. Rules for dialogue processing are divided into two groups, general rules for the … More dialogue flow and specific rules depending on the topics and tasks. Each of the topic dependent rules was constructed based on each of dialogue samples.3. Based on the above results, a system was developed for dialogue processing and its operation was checked. As the results, it was shown that, by introducing topic dependent rules, response generation is possible only with the information on the dependent words in the input speech, and that the ellipsis and the anaphora in the input speech can be compensated rather easily by referring to the topic dependent rules previously used. The advantage of the topic dependent rules was also shown to be the easiness in the addition and modification of the rules.4. A method was developed for generating surface sentences from the response content obtained by the dialogue processing. The response content is represented in the list form of case structure representation and the surface sentence is generated by adding particles and conjugating the predicates. During this process for sentence generation, phone and prosodic symbols are also generated using the linguistic information.5. Prosodic features were analyzed for the pseudo-dialogue speech. Based on the results, prosodic rules for the reading speech previously developed were modified to produce prosodic rules for the dialogue speech. These rules include the focus control.6. By combining the above results with the speech synthesis system already developed, a system was construct- ed for synthesizing speech response. Hearing test was conducted and the validity of the system was proved. A speech synthesizer was also developed to increase the speech quality. Less