语音识别音字转换中的快速容错算法  

A Fast Error-tolerant Algorithm in Decoding Module of Speech Recognition

作　　者：李明琴[1] 王作英[1] 陆大■ 

机构地区：[1]清华大学电子工程系,北京100084

出　　处：《中文信息学报》2002年第5期38-43,共6页Journal of Chinese Information Processing

基　　金：国家863高技术项目(863-306-ZD03-02-1);985重大项目(985校-22-攻关-06)

摘　　要：本文研究了汉语连续语音识别音字转换中的容错算法,以纠正声学识别的替代、插入、删除错误。为了解决容错算法的计算量问题,本文提出了两种快速算法。一是针对单独出现错误的快速容错算法;二是针对关键词的快速容错算法。快速算法有效地限制了容错算法的搜索空间,提高了计算效率。快速容错算法应用在电话对话系统中,字正确率从78.97％提高到86.68％,关键词检测正确率从80.56％提高到88.52％,并且算法运算时间满足实时性要求。An error-tolerant algorithm in decoding module of Mandarin continuous speech recognition is examined to correct substitution, insertion and deletion errors in acoustic recognition. In order to reduce the computation complexity of the error-tolerant algorithm, two fast algorithms are proposed. In both these algorithms searching space is effectively reduced, and computation efficiency is greatly improved. In the first algorithm only single isolated errors are corrected,and in the other algorithm only errors in keywords are concerned. Experiments carried out in a telephone dialogue system show that by applying the error-tolerant algorithms, word correct rate rises from 78. 97% to 86. 68% , keyword spotting rate rises from 80.56 % to 88.52% , and runtime of algorithm meets the requirement of realtime application.

关 键 词：语音识别 音字转换 快速容错算法 稳健语音识别 口语对话系统 关键词检测 搜索空间 

分 类 号：TN912.34[电子电信—通信与信息系统]