一种基于SVM的加权似然比融合算法  

Weighted likelihood ratio fusion algorithm based on SVM

作　　者：魏振华[1] 江雪峰[1] 

机构地区：[1]华北电力大学控制与计算机工程学院,北京102206

出　　处：《控制与决策》2011年第5期777-780,共4页Control and Decision

基　　金：国家自然科学基金项目(60775058);哈尔滨市科学研究基金项目(2005AXJ052);华北电力大学博士科研基金项目(200722019)

摘　　要：提出一种基于改进竞争模型加权似然比融合的方法来进行关键词的语音确认,并将误差函数用于SVM对加权向量和阈值进行优化.通过构建两个基线系统,将不同竞争模型的加权方式的比较和3种模型的比较作了仿真实验.所得结果显示,相对于反词模型和常规竞争模型,加权似然比融合方法得到的似然比对关键词识别的平均错误率分别降低了26.53%和7.78%.An approach of verification of key words the weighted likelihood ratio（LR） fusion based on the revised rival-model is proposed to be applied to the utterance.Error function is proposed in order to realize the parameter optimization based on the SVM.In the experiment stage,two baseline systems are setup.In the first experiment,the different weight methods based on the rival model are distinguished.In the second experiment,three different models are compared.The experiment result shows that,the LR computed by the weighted likelihood ratio fusion makes a better performance on the keyword spotting,which decreases 26.53% and 7.78% comparing to the normal rival-model and classical anti-model respectively.

关 键 词：置信度测量 竞争模型 语音确认 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]