摘要
提出一种新的基于近邻竞争模型的鲁棒语音确认方法。该方法通过引入目标模型的鲁棒近邻信息以提高在环境不匹配情况下的似然比确认性能。实验结果表明,在纯净语音环境下,该方法的性能与似然比方法相当,在高斯白噪声环境下,与似然比、在线垃圾模型方法相比,该方法的错误率分别下降1.2%和4.2%,在其他噪声环境下,该方法也能获得较好的确认效果。
This paper proposes a novel verification method based on the neighborhood competing models, The method can improve the performance of likelihood ratio verification method in mismatch conditions by incorporating robust neighborhood information for a target model. Compared with traditional Likelihood Ratio Test（LRT） and On-line Garbage（OLG） model verification methods, experiments show the method can provide a comparable performance to the original LRT method in clean environment, but outperform the other approaches in the Gaussian white noise condition with Equal-Error-Rate（EER） reduced by 1.2% and 4.2%, and it also achieves better verification performance in other noise conditions.
出处
《计算机工程》
CAS
CSCD
北大核心
2008年第16期4-6,共3页
Computer Engineering
基金
国家“863”计划基金资助重点项目“多语言语音识别关键技术研究与应用产品开发”(2006AA010102)
河北省教育厅基金资助项目“智能电话转接系统”(2005340)
石家庄经济学院基金资助项目“中文语音关键词检测技术关键算法的研究”(200746)
河北省科技厅基金资助项目“基于网络流媒体广告的音频内容检索技术的研究”(052135147)
关键词
语音识别
语音确认
似然比检验
近邻信息
speech recognition
utterance verification
Likelihood Ratio Test（LRT）
neighborhood information
