基于改进PNCC特征和两步区分性训练的录音设备识别方法  

A Recording Device Identification Algorithm Based on Improved PNCC Feature and Two-Step Discriminative Training

作　　者：贺前华[1] 王志锋[1,2] Alexander I Rudnicky 朱铮宇[1] 李新超[1] 

机构地区：[1]华南理工大学电子与信息学院,广东广州510640 [2]卡内基梅隆大学计算机学院,美国匹兹堡15213

出　　处：《电子学报》2014年第1期191-198,共8页Acta Electronica Sinica

基　　金：国家自然科学基金(No.60972132;No.61101160);广东省自然科学基金(No.9351064101000003;No.10451064101004651)

摘　　要：录音设备来源识别是通过分析已获取的数字语音信号从而确定其录制设备的一种技术,属于数字音频盲取证.本文提出了一种基于改进PNCC特征和两步区分性训练的录音设备识别方法,由于音频中的静音包含了完整的设备信息,且不受说话人和文本等因素的影响,因此从静音段提取改进的PNCC特征,利用了PNCC的长时帧分析去除背景噪声对设备信息的影响.在模型方面,以GMM-UBM为基准模型,并通过两步区分性训练调整集内设备模型和通用背景模型,提升模型区分能力.该方法对于30种设备闭集识别的平均正确识别率为90.23%;对于15个集内和15个集外设备的测试,等错误率为15.17%,集内平均正确识别率为96.65%,验证了本文算法的有效性.Recording device identification is a kind of blind digital audio forensic technique, which extracts digital evidence of device mechanism involved in the generation of the speech recording by analyzing the acoustic signal. This paper proposes a recording device identification algorithm which is based on improved PNCC feature and two-step discriminative training.Due to the fact that silence periods contain the device information and is not affected by speaker and texture factors, this paper extracts improved PNCC from silence periods, which uses long term analysis to remove the effect of background noise. GMM-UBM is set as the baseline system, which is improved by two steps discriminative training. The experimental result indicates that the average accuracy of recording device identification on 30 devices is 90.23% ;for 15 inset and 15 outset devices testing,the EER is 15.17% and ACC is 96.65 %, which proves the effectiveness of the proposed algorithm.

关 键 词：数字音频取证 录音设备识别 GMM-UBM 区分性训练 PNCC 

分 类 号：TN912.3[电子电信—通信与信息系统]