Large Bocabularly Spoken Word Recognition system Using Phonemic Segmentation Units

使用音素分割单元的大型口语单词识别系统

基本信息

批准号：
05555102
负责人：
IMAI Satoshi
金额：
$ 6.34万
依托单位：
Tokyo Institute of Techology
依托单位国家：
日本
项目类别：
Grant-in-Aid for Developmental Scientific Research (B)
财政年份：
1993
资助国家：
日本
起止时间：
1993 至 1994
项目状态：
已结题

来源：
https://kaken.nii.ac.jp/grant/KAKENHI-PROJECT-05555102/
关键词：
Spoken word Recognition Phonemic Segmentation Phoneme Labeling Large Vocabulary Multiple Referenc Pattern Parallel Phoneme Labeling Method Segment Lattice Multiple Segmentation Method 大語彙化

项目摘要

In this research project, we substantiated that the speech recognition method based on the phonemic segmentation and phoneme labeling was very effective for the large vocabulary spoken word recognition, and we developed a high performance large vocabulary spoken word recognition system using the phonemic segmentation units.This spoken word recognition system is composed of the following subsystems : an acoustic analysis subsystem, phonemic segmentation units, phoneme labeling subsystem and word matching subsystem. one ofproblem of this word recognition system errors in the phonemic segmentation and phoneme labeling. We tried to improve the system in the phonemic segmentation and phoneme labeling.Trough this research project, we got the following good results.(1) We realized a high performance automatic phonemic segmentation unit for speaker and context independent Japanese speech recognition system. We substantiated that this segmentation unit was effective for the large vocabulary word recognition.(2) We developed a higher performance large vocabulary spoken word recognition system using the phonemic segmentation unit and phoneme labeling system.Experiments were carried out using the dictionaries of 1845 words and 4915 words to evaluate the system. The word recognition rates for the first candidate were found to be 96.5% and 94.5% for 1845 word and 4915 word dictionaries respectively. An estimated recognition rate for 20000 word dictionary was approximately 90%.(3) We proposed the parallel phonemic segmentation method in order to achieve a higher word recognition rate. Using the parallel phonemic segmentation unit, we obtained 1 or 2% higher recognition rate for 4915 word dictionay. We also proposed the parallel phoneme labeling method, and substantiated the method is very effective for realizing a higher recognition rate.

在本研究项目中，我们证实了基于音素分割和音素标记的语音识别方法对于大词汇量口语单词识别非常有效，并且我们开发了一种使用音素分割单元的高性能大词汇量口语单词识别系统。口语单词识别系统由以下子系统组成：声学分析子系统、音素分割单元、音素标注子系统和单词匹配子系统。该单词识别系统的问题之一是在音素分割和音素标注方面存在错误。我们试图在音素分割和音素标记方面改进该系统。通过这个研究项目，我们得到了以下良好的成果：(1)我们实现了一种用于说话人和上下文无关的日语语音识别系统的高性能自动音素分割单元。我们证实了该切分单元对于大词汇量单词识别是有效的。(2)利用音素切分单元和音素标记系统开发了性能更高的大词汇量口语单词识别系统。使用1845个单词的词典和1845个单词的词典进行了实验。 4915字评价系统。对于 1845 个单词和 4915 个单词的词典，第一个候选的单词识别率分别为 96.5% 和 94.5%。对20000个单词的词典的估计识别率约为90%。(3)我们提出了并行音素分割方法以获得更高的单词识别率。使用并行音素分割单元，我们对 4915 个单词的词典的识别率提高了 1% 或 2%。我们还提出了并行音素标记方法，并证实该方法对于实现更高的识别率非常有效。