Knowledge Base and Its Environment for Continuous Speech Recongnition

本研究的主要目的是构建一个连续语音识别的知识库及其支持环境。该系统模拟人类专家的行为，该专家可以读取共振峰频率和功率等特征参数的轨迹。专家知识以产生式规则的形式嵌入知识库中。该环境由语音库、规则库、规则翻译器和接口模块四部分组成。语音数据库存储带有右段标签的语音数据，并根据请求提供合适的语音数据。规则数据库存储了一些规则文档，如作者姓名和执行历史，其中提供了指向语音数据库中数据的指针。它还管理规则的版本，这在规则修改中起着重要作用。规则翻译器将我们设计的高级规则语言编写的规则翻译成OPS5代码。接口模块将上述三个组件集成到一个统一的系统中，实现了物理交互，对6名男性乌达特人发出的5种连续语音，在给定正确音素边界的情况下，分割率和识别率分别达到了95%和90%。该环境将规则开发时间减少到不到以前的十分之一。此外，在保持识别率的前提下，利用环境对大部分规则进行了重新组织，使规则的数量减少到原来知识库的70%。这些研究成果为建立一种新的基于知识工程技术的语音识别方法奠定了基础。

The main objective of our research is to construct a knowledge base for continuous speech recognition and its support environment. The system simulates the behavior of a human expert who can read the trajectories of feature parameters such as formant frequencies and power. The expertise is embedded in the knowledge base in the form of production rules. The environment is composed of four components such as speech database, rule database, rule translator and interface module. Speech database stores speech data with right segment labels and gives appropriate ones on request. Rule database stores some documents of rules such as name of the author and execution history which provides with pointers to data in speech database. It also manages versions of rules, which plays an important role in rule modification. Rule translator translates rules written in a highlevel rule language designed by us into OPS5 codes. Interface module integrates the above three components into a unified system and enables sophysticated interactions.We obtained segmentation rate of 95% and recognition rate of 90% given the correct phoneme boundaries for five kinds of continuous speech uttered by six male udalts. The environment reduces the rule development time to less than one tenth of the previous one. Furtheremore, most of the rules were reorganized with the aid of the environment and the number of the rules were recuced to 70% of the original knowledge base retaining the recognition rate. All the results of the research contribute to the first step to establishment of a new methodology of speech recognition basd on knowledge engineering techniques.