融合MobileNet与GhostVLAD的欺骗语音检测  

Spoofed Speech Detection Based on MobileNet and GhostVLAD

作　　者：闫佳 冯爽[1,2,3] YAN Jia;FENG Shuang(School of Computer and Cyber Sciences,Communication University of China;Key Laboratory of Convergent Media and Intelligent Technology;State Key Laboratory of Media Convergence and Communication,Beijing 100024,China)

机构地区：[1]中国传媒大学计算机与网络空间安全学院 [2]智能融媒体教育部重点实验室 [3]媒体融合与传播国家重点实验室,北京100024

出　　处：《软件导刊》2021年第3期204-208,共5页Software Guide

摘　　要：欺骗语音已经对自动说话人识别系统产生严重影响。语音转换、语音合成和语音重放是对ASV系统进行欺骗的3种主要手段,但流行的欺骗检测方法使用的卷积神经网络大多层数较深、网络较复杂,不适合在移动设备以及嵌入式设备上运行。针对这一问题,提出一种适用于3种欺骗情况下的基于轻量型神经网络Mo⁃bileNet和GhostVLAD(Ghost Vector of Local Aggregated Descriptors)方法相结合的算法。首先对语音数据进行增广并提取常数Q倒谱系数和振幅频谱图,将其作为输入特征;然后将MobileNetV2或V3-large网络的最后一个池化层替换为GhostVLAD聚合层;最后使用端到端的优化方法对真实语音和欺骗语音进行识别。在ASVspoof 2019数据集上进行实验,结果表明该算法效果较好,相比基线系统在等错误率上分别降低了38%和13%。Spoofed speech has seriously affected the automatic speaker recognition system.Speech conversion,speech synthesis and speech replay are three main methods to spoof ASV system.However,most of the convolutional neural networks used by popular detec⁃tion methods are deep and complex,which are not suitable for mobile devices and embedded devices.To solve this problem,this paper proposes an algorithm based on the combination of the lightweight neural network—MobileNet and GhostVLAD,which can be used in three cases.Firstly,the speech data is augmented.Then,constant Q-cepstrum coefficient and amplitude spectrum are extracted as in⁃put features;secondly,the last pooling layer of MobileNetV2(or V3 large)network is replaced by GhostVLAD aggregation layer.Fi⁃nally,the end-to-end optimization method is used to recognize real speech and spoofed speech.The experimental results on ASVspoof 2019 dataset show that the proposed algorithm achieves better results,which reduces the equal error rate by 38%and 13%respectively compared with the baseline system.

关 键 词：欺骗语音检测 ASV spoof MobileNet GhostVLAD 

分 类 号：TP319[自动化与计算机技术—计算机软件与理论]