VoIP压缩码流说话人识别研究.pdfVIP

2009 年4 月 第7 期 计舰工程 目 35 嘟 协1.35 No.7 Computer Engineering April 2009 中回分自能号I TN912.3 文献标识钩IA 文2能偏号I 1脚一.3428(2剧创07….()180-03 ·人工智能及识别技术· VoIP 压缩码流说话人识别研究 庸晖，学粥穗，周丹，张连梅 (解放军信息工模大学信息工程学院，郑州 45以用2) 摘要t 研究惑子微蕉类算法的 VolP Æ缩码流说话人识别算法。纷出或接从 0729， 0723.1 (6.3 Kbl时， 0723.1(5.3 Kb!s)压缩说裔的硝流中 提取识别参数，以微聚3能算l!作为识别结构的说汤，人识别算法。实验结果发明，对比在压缩码流中使用同样识别参数的OMM模型，微襄 樊算法在识别正确率和效率上都布很大的提底。 关键词:说话，人识别;微聚类;压缩参数 Research 00 Speaker Recogoitioo from Compressed VoIP Packet Stream TANG Hui , LI Bi-cheng, QU Dan, ZHANG Lian-hai (Institu侃。f Infonnation Engineering , PLA Informatíon Engineering Uníversity, Zhengzhou 45似阳) (Abstract )This paper presenωcompres优d speaker recognítion approach for VoIP(Voice over 1P) which is based on the micro-clustering algorithm. It designs a fu制nework based on the micro-clustering algoríthm and performs speaker reωgnition on the feaωre vector whích is directly extracted from 0729, 0723.1(6.3 Kb/s), 0723ι1(5.3 Kb/s) compresωd stre似n. Experímentai result 恼。ws that the new me由od is more accura优 and efficient than the widely used Oaussian Mixture Model(OMM) which uses the same feaωre vector. IKey wor叫 speaker reco伊白tion; micro咄lsterìng; compres则parameter 1 式是 0.723.1和 0.729 算怯，本文研究了这2 种压缩码梳的特 概述 随辛苦 VoIP(Voice over IP)业务的持续增长，越来蜡需要针