基于动态点阵匹配算法的二阶关键词识别  

Dynamic match phone-lattice algorithm for second order keyword spotting

作　　者：刘晓明[1] 冯晓荣[1] 班超帆[1] 

机构地区：[1]重庆大学通信工程学院,重庆400030

出　　处：《吉林大学学报（工学版）》2012年第3期771-775,共5页Journal of Jilin University:Engineering and Technology Edition

基　　金：国家自然科学基金项目(51035008)

摘　　要：为了避免关键词权值估计错误,提出了基于上下文相关最大后验概率准则(CDMAP)的改进二阶KWS模型,有效地解决了关键词权值对参数估计的负面影响。同时,采用动态点阵匹配算法(DMPLS)进行错误补偿,解决了由连续语音识别器产生的插入、删除和替换错误而导致识别准确率下降的问题,提高了系统的检出率和鲁棒性。实验结果表明,该模型不仅在较低误警率的条件下检出率得到很大的改善,而且系统鲁棒性也有明显的增强。In order to solve the problem of estimation error of Keyword weight,an improved model of second KWS is proposed.This model is based on the Context-Dependent Maximum a Posteriori Probability（CDMAP） method.Meanwhile,a dynamic match phone-lattice algorithm is introduced.This algorithm can efficiently solve the problem of inherent phone recognition error,which adversely affects the recognition accuracy.Experiment results show that the system performance is greatly enhanced by the dynamic match phone-lattice algorithm for both robustness and detection accuracy.

关 键 词：通信技术 关键词权值 动态点阵匹配 关键词检出 

分 类 号：TN912.34[电子电信—通信与信息系统]