Japanese text dictation system for official reports

官方报告的日语文本听写系统

基本信息

批准号：
07558042
负责人：
MAKINO Shozo
金额：
$ 3.39万
依托单位：
Tohoku University
依托单位国家：
日本
项目类别：
Grant-in-Aid for Scientific Research (B)
财政年份：
1995
资助国家：
日本
起止时间：
1995 至 1997
项目状态：
已结题

项目摘要

The experts often make official reports such as for estimation of real estiate, for medico-legal autopsy and so on. It is a time-consuming job to make official reports. If speech input is automatically transformed to sentenses, the load of making document will be decreased. In current speech input system, it is still difficult to automatically recognize continuously spoken general sentences. But, when the condition is so limited that a user is special, 1) a user is limited to a specified speaker, 2) the expression of document is almost decided, and the structure of sentence is comparatively simple, 3) the vocabulary number in the official reports is 4000 from 3000, and the vocabulary number decrease more by deviding the document into each of parts. These conditions will lighten load to a device, and the device will be developed with ease.In this research, we have developed a sentense recognition system for autopsy reports. Firstly, we built an automaton by ECGI method to represent the structure of sentence. Then we defined the distance between the states of an automaton to strengthen correspondence to the words, the appearance of which was expected. Based on this definition, we developed a method in which an automaton was revised and generalized. For phoneme recognition, we used the model sound method developed by Niyada. We, all the members, put the above-mentioned methods together and made the sound input system of autopsy findings. The system ran to recognize sounds without time delay. Since the precision of sound recognition is not enough, the improvement of the system will be continued in future.

专家经常出具房产估价、法医学尸检等正式报告。制作正式报告是一项耗时的工作。如果将语音输入自动转换为句子，则可以减少制作文档的负担。在当前的语音输入系统中，自动识别连续说出的一般句子仍然很困难。但是，当条件有限，用户比较特殊时，1）用户仅限于特定的说话人，2）文档的表达方式基本确定，句子的结构比较简单，3）文档中的词汇量官方报告从3000个变成了4000个，通过将文档分成各个部分，词汇量减少得更多。这些条件将减轻设备的负载，并且设备的开发将变得容易。在本研究中，我们开发了尸检报告的句子识别系统。首先，我们通过ECGI方法构建了一个自动机来表示句子的结构。然后，我们定义了自动机状态之间的距离，以加强与预期出现的单词的对应关系。基于这个定义，我们开发了一种对自动机进行修改和推广的方法。对于音素识别，我们使用了 Niyada 开发的模型声音方法。我们全体成员将上述方法综合起来，制定了完善的尸检结果输入系统。系统运行来识别声音，没有时间延迟。由于声音识别的精度还不够，未来还将继续对系统进行改进。