Development of Speech Understanding System Using Artificial Intelligence Techniques

本研究的目标包括建立一个连续的韩语语音识别系统，以及对基于知识的技术及其在语音理解系统中的应用进行基础研究。理解连续语音通常是一个棘手的问题，因为声学信息是不可靠的。这就要求我们不仅要使用声学信息，还要使用语言信息来消除声学处理的歧义。在这项与韩国研究者的国际研究中，我们频繁地讨论了常见的问题，并获得了我们的语音识别专家系统SPREX应用于韩语语音识别的可能性。以下是对今年取得的研究成果的总结。1)我们调查了韩语的语音特征，发现了几个不同之处：它的元音包括半元音和许多鼻音大约多两倍。因此，我们扩展了SPREX的机制和知识，使其适用于韩语语音识别。此外，我们发现SPREX中负责描述特征参数轨迹的前端模块应该重新编程以应对韩语语音描述。2)为了准备语音理解系统的构建，我们还调查了韩语的特点，以了解日语与韩语的区别。通过调查，我们发现语音理解系统的语言处理子系统的设计没有太大的差异，即我们目前为日语开发的系统原则上适用于韩语。3)语音数据库在语音研究中起着重要的作用，特别是在不同语言的研究中。虽然还处于初级阶段，但我们已经开始了对韩语和日语语音数据库结构的研究。

The objectives of this research include to build a continuous speech recognition system for Korean speech as well as performing fundamental research for knowledge-based techniques and its applications to speech understanding system. Understanding of continuous speech is generally a tough problem, since acoustic information is unreliable. This requires us to employ not only acoustic information but also linguistic one to disambiguate the acoustic processing. In this international research with a Korean researcher, we discussed the common issues frequently and obtained the possibility of the application of our speech recognition expert system named SPREX to Korean speech recognition. The following are the summary of the research results obtained this year.1) We investigated the characteristics of Korean speech and found several differences : It has about two-times more vowels including semi-vowels and many nasals. So, we extended the mechanism and the knowledge of the SPREX to enable it applicable to Korean speech recognition. Furthermore, we found the front end module in SPREX responsible for describing the trajectory of feature parameters should re-program in order to cope with Korean speech description.2) For preparation to the speech understanding system building, we also investigated the characteristics of Korea language to know the difference between Japanese and it. Based on the investigation, we found no serious difference in designing the language processing subsystem for the speech understanding system, that is, the system we currently develop for Japanese language is applicable to Korean in principle.3) Speech data base plays an important role in speech research, especially in our research on different languages. Although it is just in the elementary stage, we began the investigation of the structure of speech data base for Korean spoken language as well as Japanese.