连续语音关键词识别系统的拒识方法研究  

Research on rejection method for continuous speech

作　　者：徐明星[1] 郑方[1] 吴文虎[1] 方棣棠[1] 

机构地区：[1]清华大学计算机科学与技术系

出　　处：《清华大学学报（自然科学版）》1998年第S1期92-94,共3页Journal of Tsinghua University(Science and Technology)

基　　金：国家"八六三"高技术项目

摘　　要：为解决在连续语音关键词识别系统中，对语音识别候选进行拒识问题，采用中心距离连续概率模型，提出了以概率统计理论和试验分析为基础的新拒识方法：“特征临界区域百分比”方法和“识别分数间隙”方法。这两种方法可以联合使用。为了考察拒识方法的拒识性能，定义了４个评价参数，并在一个来自现实世界的语音大数据库上，测试了２万个样本。拒识的平均正识率为８６．３３％，拒识后平均候选个数降为３．４６（未进行拒识前是１０名候选），总的拒识错误率为０．２７％。试验结果表明，本文提出的两种拒识方法是有效的和可行的。In order to reject speech recognition candidates especially in the KeyWords spotting system, a new rejection method based on center distance continues probability model (CDCPM) is proposed, named CAP, which is feature percentage in critical area (CAP) according to the probability theories. Also another rejection method named recognition score gap (RSG) is proposed to cooperate with CAP. Experiments are done across a large real world database with 20 000 test samples. The average recognition accuracy is 86.33% with 3.46 candidate number (the number was 10 before rejection) on an average, and the total error rejection rate is 0.27%. The results show that the rejection methods are feasible and effective.

关 键 词：语音识别 拒识 中心距离连续概率模型(CDCPM) 特征临界区域百分比(CAP) 识别分数间隙(GSP) 

分 类 号：TP391.42[自动化与计算机技术—计算机应用技术]