Development of a speech understanding system

本研究的目标是开发理解口语对话所需的基本技术，包括基于知识的语音识别系统，自然语言处理中的非单调推理和对话建模。本文的主要研究成果如下：1）验证了基于知识的韩语语音识别方法的有效性。此外，还提出了一些改进语音识别的新思想。为了避免分割困难，引入了非均匀单元。每个单元在其两端都有其静止点，中间有其瞬态部分。参数轨迹由符号表示和模糊语言变量描述。在后处理部分，对语音数据进行冗余处理，以提高识别系统的性能。原型系统对未知长度的连续韩语数字语音进行了测试，识别率达到97%。2）由于声学信息的不可靠性，连续语音的理解通常是一个坚韧。由于歧义信息的组合非常大，因此有效的搜索机制是必不可少的。在此基础上，提出了一个基于ATMS的语音理解系统框架。ATMS是一种非单调推理方法。ATMS的引入将自然语言处理的时间从64秒减少到45秒，用于理解8个日语单词的语音。3）提出了两种描述对话结构的对话模型，用于理解口语对话。一种是描述由刺激和反应组成的话语对的SR计划模型。另一种是主题分组网络（TPN），与语篇片段相对应。基于这些对话模型开发了一个预测下一个话语的机制，并在一些样本对话上进行了评估。

The objective of this research is development of fundamental techniques necessary to understanding spoken dialogue, which include knowledge-based speech recognition system, non-monotonic reasoning in natural language processing, and dialogue modeling. The following are the summary of the research results.1) We verified the efficiency of the knowledge-based approach for Korean speech recognition. Furthermore, some new ideas were proposed to improve the speech recognition. To avoid the difficulties in segmentation, a non-uniform unit is introduced. Every unit has its stationary point at each end of the unit, and transient part in the middle. The parameter trajectory is described by symbolic representation and fuzzy linguistic variables. Redundancy of speech data is used to improve the performance of the recognition system in the post-processor. The prototype system was tested with continuous Korean digit speech of unknown length, and the recognition rate of 97% was obtained.2) Understanding of continuous speech is generally a tough problem, since acoustic information is unreliable. An efficient search mechanism is indispensable because the combination of ambiguous information is very large. Then, we developed a framework of speech understanding system based on ATMS, which is a method of non-monotonic reasoning. The introduction of ATMS reduced elapsed time of natural language processing from 64 sec to 45 sec for understanding speech of 8 Japanese sentences.3) Two kinds of dialogue model characterizing structures in dialogue were proposed for understanding spoken dialogue. One is the SR-plan model which describes utterance pairs composed of the stimulus and the response. The other is Topic Packet Network (TPN) and corresponds to the discourse segments. A mechanism for predicting the next utterance was also developed based on these dialogue models and evaluated on some sample dialogues.