改进的非匹配信道法庭自动说话人识别方法  

Improved Method of Mismatched Channel Forensic Automated Speaker Recognition

作　　者：王华朋[1,2] 杨军[1] 吴鸣[1] 许勇[1] 

机构地区：[1]中国科学院噪声与振动重点实验室(声学研究所),北京100190 [2]中国刑事警察学院,辽宁沈阳110854

出　　处：《电声技术》2013年第4期61-67,71,共8页Audio Engineering

基　　金：国家自然科学基金项目(11004217;11074279)

摘　　要：在新的似然比法庭证据评估体系下,为降低对问题语音和嫌疑人语音样本数量的要求,提出了改进的基于GMM-UBM识别核心的法庭自动说话人识别系统。为降低信道差异对法庭自动说话人识别系统性能的影响,提出了一种不同信道间说话人模型补偿与Logistic回归混合处理的方法。固定电话信道和移动电话信道的实验结果表明:改进的自动识别系统在非匹配信道中使用4秒的测试语音,与未使用模型补偿的系统相比,错误认定率从28.95%下降为6.32%,大幅度提高了系统在非匹配条件下的可靠性。A modified forensic automatic speaker recognition （FASR） system based on GMM - UBM is proposed to re- duce the demand for the quantity of suspect＇ s speech and questioned speech, in the new forensic evidence evaluation framework of likelihood ratio. Furthermore, a new method which combines speaker model compensating algorithm and lo- gistic regression is proposed to degrade the influence of mismatched transmission channel. The mismatched channels are of fixed telephone network and global system for mobile communications （GSM） transmission. Experiment results show that, compared with the FASR system without model compensation, the false positive rate of the modified system with the four seconds test utterance degrades from 28.95% to 6.32% , which greatly enhances the reliability of FASR system in mis- matched channel condition.

关 键 词：非匹配信道 法庭说话人识别 似然比 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]