基于卷积神经网络的电子变调语音检测算法

doi:10.11959/j.issn.1000-0801.2018041

Abstract

Abstract:

An electronic disguised voice detection algorithm based on the statistical features of MFCC and the convolution neural network was proposed.Firstly,the statistical features of MFCC were extracted and reconstructed as the input of convolution neural network.Considering the convolution kernel size,the number of convolution kernels and the pooling size,24 different network structures were evaluated in this work.Finally,the convolution neural network structure which could be effectively used for electronic disguised voice detection was determined.The experimental results show that the proposed algorithm can effectively detect the trace of electronic disguising.Meanwhile,the specific forgery operation of the electronic disguised voice can also be estimated.

Key words: electronic disguised voice, convolutional neural network, MFCC

CLC Number:

TP391

Hongwei XU,Diqun YAN,Fan YANG,Rangding WANG,Chao JIN,Li XIANG. Detection algorithm of electronic disguised voice based on convolutional neural network[J]. Telecommunications Science, 2018, 34(2): 46-57.

Figures/Tables 9

网络名称	训练方法			测试方法分类结果A₁/A₂				FAR
网络名称	训练方法	Audacity		CoolEdit		RTIS		FAR
a-1	Audacity	98.87%	99.35%	98.77%	99.19%	98.67%	98.77%	2.08%
	CoolEdit	94.59%	93.12%	98.92%	99.62%	97.74%	97.38%	2.46%
	RTISI	97.37%	97.20%	98.21%	99.32%	98.97%	98.42%	1.84%
a-2	Audacity	99.02%	99.31%	98.26%	98.15%	97.41%	97.00%	1.54%
	CoolEdit	95.03%	94.12%	98.74%	99.69%	97.89%	97.77%	3.15%
	RTISI	97.23%	97.11%	98.67%	99.07%	98.51%	98.21%	1.92%
b-1	Audacity	99.28%	99.42%	98.92%	98.88%	97.56%	97.30%	1.00%
	CoolEdit	95.46%	94.35%	99.10%	99.81%	97.46%	97.00%	2.31%
	RTISI	97.92%	98.96%	98.20%	98.08%	98.31%	97.54%	1.15%
b-2	Audacity	99.26%	99.38%	98.87%	98.81%	98.64%	98.46%	1.00%
	CoolEdit	96.31%	96.69%	98.85%	99.50%	98.23%	98.07%	2.46%
	RTISI	96.56%	96.04%	97.87%	97.11%	99.36%	99.23%	0.92%
b-3	Audacity	99.54%	99.77%	98.87%	98.77%	98.33%	98.62%	0.92%
	CoolEdit	97.10%	96.46%	99.36%	99.85%	98.84%	98.80%	1.62%
	RTISI	99.23%	99.18%	99.69%	99.57%	99.26%	99.83%	0.85%
c-1	Audacity	99.41%	99.69%	99.00%	99.03%	98.59%	99.58%	1.08%
	CoolEdit	95.05%	93.46%	99.13%	99.58%	98.26%	98.38%	1.77%
	RTISI	97.18%	95.13%	97.51%	96.62%	98.87%	98.73%	1.00%
c-2	Audacity	99.51%	99.77%	98.91%	99.00%	98.92%	98.92%	0.92%
	CoolEdit	96.03%	94.69%	99.41%	99.77%	98.38%	98.15%	1.31%
	RTISI	96.54%	96.15%	97.51%	96.58%	99.18%	99.04%	0.63%
c-3	Audacity	99.51%	99.81%	99.00%	99.04%	98.64%	98.54%	1.08%
	CoolEdit	96.38%	95.12%	99.51%	99.81%	98.36%	98.19%	1.08%
	RTISI	97.54%	96.00%	98.03%	97.27%	99.46%	99.35%	1.31%
c-4	Audacity	99.28%	99.65%	98.90%	99.08%	98.46%	98.27%	1.46%
	CoolEdit	95.67%	94.58%	99.21%	99.88%	98.00%	98.12%	2.15%
	RTISI	96.41%	95.91%	97.87%	97.00%	99.05%	98.89%	0.62%
c-5	Audacity	99.49%	99.58%	98.79%	98.54%	98.59%	98.35%	0.69%
	CoolEdit	96.36%	95.08%	99.44%	99.69%	98.21%	97.69%	1.08%
	RTISI	97.05%	96.77%	997.97%	97.42%	99.36%	99.23%	0.78%
c-6	Audacity	99.49%	99.81%	98.95%	99.00%	98.67%	98.69%	1.15%
	CoolEdit	96.62%	95.64%	99.44%	99.88%	98.41%	98.23%	1.46%
	RTISI	97.59%	96.12%	97.89%	97.23%	99.23%	99.04%	0.78%
d-1	Audacity	99.51%	99.77%	99.13%	99.19%	98.26%	98.27%	1.00%
	CoolEdit	95.21%	93.62%	99.28%	99.73%	98.43%	98.27%	1.62%
	RTISI	96.53%	96.63%	97.87%	97.07%	99.38%	99.42%	1.09%
d-2	Audacity	99.36%	99.62%	98.69%	98.62%	98.48%	98.23%	1.15%
	CoolEdit	96.59%	95.42%	99.49%	99.77%	98.23%	98.04%	1.08%
	RTISI	97.77%	97.26%	97.54%	96.19%	99.13%	99.04%	1.09%
d-3	Audacity	99.38%	99.58%	99.00%	98.96%	98.41%	98.27%	0.92%
	CoolEdit	95.85%	94.62%	99.31%	99.81%	98.10%	97.92%	1.69%
	RTISI	96.67%	97.59%	98.35%	96.58%	99.21%	99.12%	0.61%
d-4	Audacity	99.44%	99.58%	98.90%	98.77%	98.67%	98.38%	0.85%
	CoolEdit	96.36%	95.08%	99.51%	99.81%	98.26%	97.96%	1.08%
	RTISI	97.97%	97.26%	97.77%	96.08%	99.44%	99.35%	0.78%
d-5	Audacity	99.36%	99.81%	98.97%	99.23%	98.69%	98.65%	1.54%
	CoolEdit	96.23%	94.92%	99.49%	99.81%	98.31%	98.19%	1.15%
	RTISI	97.59%	97.72%	97.11%	96.85%	99.15%	98.96%	0.85%
e-1	Audacity	99.26%	99.73%	99.00%	99.31%	98.49%	98.27%	1.62%
	CoolEdit	95.46%	93.88%	99.28%	99.62%	98.21%	97.85%	1.38%
	RTISI	97.08%	97.59%	98.54%	96.69%	99.15%	99.00%	0.64%
e-2	Audacity	99.46%	99.62%	98.87%	98.69%	98.41%	98.54%	0.77%
	CoolEdit	95.87%	94.35%	99.43%	99.69%	98.03%	98.46%	1.08%
	RTISI	98.53%	98.28%	98.08%	97.65%	99.15%	98.96%	0.78%
e-3	Audacity	99.44%	99.77%	99.10%	99.27%	98.48%	98.35%	1.23%
	CoolEdit	95.41%	93.96%	99.28%	99.78%	98.28%	98.35%	1.69%
	RTISI	97.89%	97.69%	97.54%	96.81%	99.31%	99.12%	0.89%
e-4	Audacity	99.33%	99.69%	99.05%	99.31%	98.64%	98.50%	1.38%
	CoolEdit	96.59%	95.73%	99.31%	99.81%	98.33%	98.04%	1.69%
	RTISI	97.85%	96.95%	96.65%	95.65%	99.38%	99.35%	0.78%
e-5	Audacity	99.46%	99.77%	99.08%	99.19%	98.56%	98.58%	1.15%
	CoolEdit	96.28%	95.50%	99.18%	99.84%	98.15%	97.73%	2.15%
	RTISI	97.05%	96.38%	97.85%	96.35%	99.31%	99.19%	0.85%
e-6	Audacity	99.41%	99.69%	98.90%	98.88%	98.69%	98.62%	1.08%
	CoolEdit	95.08%	95.08%	99.36%	99.81%	98.38%	98.08%	1.54%
	RTISI	97.15%	97.67%	97.54%	96.81%	99.36%	99.31%	0.78%
e-7	Audacity	99.49%	99.81%	99.05%	99.15%	98.31%	97.85%	1.15%
	CoolEdit	94.97%	93.46%	99.21%	99.81%	99.18%	99.50%	2.00%
	RTISI	97.23%	96.00%	98.62%	97.12%	99.27%	99.04%	0.85%
e-8	Audacity	99.49%	99.65%	99.08%	99.04%	98.28%	98.23%	0.85%
	CoolEdit	96.28%	94.92%	99.46%	99.69%	98.31%	98.23%	1.00%
	RTISI	97.13%	96.46%	97.12%	97.31%	99.33%	99.27%	0.53%

References 17

[1]	RODMAN R , . Speaker recognition of disguised voices:A program for research[C]// Consortium on Speech Technology in Conjunction with the Conference on Speaker Recognition by Man and Machine:Directions for Forensic Applications,Oct 8-11,1998, Ankara,Turkey.[S.l.:s.n] 1998: 9-22.
[2]	WU H , WANG Y , HUANG J . Blind detection of electronic disguised voice[C]// 2013 IEEE International Conference on Acoustics,Speech and Signal Processing,May 26-31,2013,Vancouver,Canada. Piscataway:IEEE Press, 2013: 3013-3017.
[3]	WU H , WANG Y , HUANG J . Identification of electronic disguised voices[J]. IEEE Transactions on Information Forensics and Security, 2014,9(3): 489-500.
[4]	CAO W , WANG H . Identification of Electronic Disguised Voices in the Noisy Environment[C]// International Workshop on Digital-forensics and Watermarking,Sep 17-19,2016, Beijing China.[S.l.:s.n] 2016: 75-87.
[5]	ZHENG F , ZHANG G , SONG Z . Comparison of different implementations of MFCC[J]. Journal of Computer science and Technology, 2001,16(6): 582-589.
[6]	KRIZHEVSKY A , SUTSKEVER I , HINTON G E . ImageNet classification with deep convolutional neural networks[C]// Advances in neural information processing systems,Dec 3-8,2012,Lake Tahoe,USA. New York:ACM Press, 2012: 1097-1105.
[7]	SIMONYAN K , ZISSERMAN A . Very deep convolutional networks for large-scale image recognition[J]. arXiv preprint arXiv:1409.1556, 2014.
[8]	ROUCOS S , WILGUS A . High quality time-scale modification for speech[C]// ICASSP’85:Proceedings of IEEE International Conference on Acoustics,Speech,and Signal Processing,Apri 26-29,1985,Florida,USA. Piscataway:IEEE Press, 1985: 493-496.
[9]	ZHU X , BEAUREGARD G , WYSE L . Real-time signal estimation from modified short-time Fourier transform magnitude spectra[J]. IEEE Transactions on Audio Speech ＆ Language Processing, 2007,15(5): 1645-1653.
[10]	Time-scale/pitch modification[EB/OL].(2009-11-24)[201709-27]. .
[11]	ZHU X , BEAUREGARD G T . Real-time signal estimation from modified short-time Fourier transform magnitude spectra[J]. IEEE Transactions on Audio Speech ＆ Language Processing, 2007,15(5): 1645-1653.
[12]	TREHUB S E , COHEN A J , THORPE L A . Development of the perception of musical relations:semitone and diatonic structure[J]. Journal of Experimental Psychology Human Perception ＆ Performance, 1986,12(3): 295.
[13]	Audacity:Free Audio Editor and Recorder[EB/OL].(2016-01-20)[2017-03-27]. .
[14]	Cool Edit Pro is Now Adobe Audition[EB/OL].(2012-11-08)[2017-03-27]. .
[15]	LECUN Y , BOTTOU L , BENGIO Y . Gradient-based learning applied to document recognition[J]. Proceedings of the IEEE, 1998,86(11): 2278-2324.
[16]	SRIVASTAVA N , HINTON G E , KRIZHEVSKY A . Dropout:a simple way to prevent neural networks from overfitting[J]. Journal of Machine Learning Research, 2014,15(1): 1929-1958.
[17]	CHOLLET F . Keras[EB/OL].（2016-09-16）[2016-11-24]. .

Metrics

Recommended 0

No Suggested Reading articles found!

网络参数名称	参数	输出数据结构
卷积	32@2×2	(2，77，32)
激活函数	Relu	(2，77，32)
卷积	32@2×2	(1，76，32)
激活函数	Relu	(1，76，32)
dropout	0.25	(1，76，32)
全连接	128	128
激活函数	Relu	128
dropout	0.25	128
softmax	-	3
激活函数	Relu	3

网络名称	实验	网络结构
a-1	卷积单元个数	卷积层:32×2×2池化:1×2全连接层:128
a-2		卷积层:32×2×2 池化:1×2 卷层:32×2×2 池化:1×2全连接层:128
b-1	池化层个数和位置	卷积层:32×2×2卷积层:32×2×2池化:1×2全连层:128
b-2		卷积层:32×2×2池化:1×2卷积层:32×2×2全连接层:128
b-3		卷积层:32×2×2 卷积层:32×2×2 全连接层:128
c-1	卷积核个数	卷积层:8×2×2卷积层:8×2×2全连接层:128
c-2		卷积层:16×2×2 卷积层:16×2×2 全连接层:128
c-3		卷积层:64×2×2 卷积层:64×2×2 全连接层:128
c-4		卷积层:8×2×2卷积层:16×2×2全连接层:128
c-5		卷积层:16×2×2 卷积层:32×2×2 全连接层:128
c-6		卷积层:32×2×2 卷积层:64×2×2 全连接层:128
d-1	全连接层神经元数	卷积层:32×2×2卷积层:32×2×2全连接层:48
d-2		卷积层:32×2×2卷积层:32×2×2全连接层:64
d-3		卷积层:32×2×2卷积层:32×2×2全连接层:96
d-4		卷积层:32×2×2 卷积层:32×2×2 全连接层:160
d-5		卷积层:32×2×2 卷积层:32×2×2 全连接层:256
e-1	卷积核形状	卷积层:32×2×2 卷积层:32×1×2 全连接层:128
e-2		卷积层:32×2×2 卷积层:32×1×3 全连接层:128
e-3		卷积层:32×2×2 卷积层:32×2×1 全连接层:128
e-4		卷积层:32×2×1 卷积层:32×2×2 全连接层:128
e-5		卷积层:32×3×1 卷积层:32×1×2 全连接层:128
e-6		卷积层:32×3×1 卷积层:32×1×3 全连接层:128
e-7		卷积层:32×3×1 卷积层:32×1×5 全连接层:128
e-8		卷积层:32×2×1 卷积层:32×1×5 全连接层:128

Detection algorithm of electronic disguised voice based on convolutional neural network

RichHTML

PDF下载

Knowledge

Abstract

Cite this article

share this article

Figures/Tables 9

References 17

Related Articles 5

Metrics

Recommended 0

分类器类型	训练方法	测试方法/检测准确率A₂			FAR
分类器类型	训练方法	Audacity	CoolEdit	RTISI	FAR
SVM-参考文献[2]	Audacity	99.83%	99.27%	97.75%	1.83%
	CoolEdit	97.29%	99.81%	98.62%	3.35%
	RTISI	98.75%	99.35%	99.80%	1.75%
CNN-本文	Audacity	99.77%	98.77%	98.62%	0.92%
	CoolEdit	96.46%	99.85%	98.80%	1.62%
	RTISI	99.18%	99.57%	99.83%	0.85%

[1]	Liang CHEN, Yaguan QIAN, Zhiqiang HE, Xiaohui GUAN, Bin WANG, Xing WANG. A flexible pruning on deep convolutional neural networks [J]. Telecommunications Science, 2022, 38(1): 83-94.
[2]	Zihao LIU, Xiaojun JIA, Sulan ZHANG, Zhiling XU, Jun ZHANG. Vibe++ background segmentation method combining MeanShift clustering analysis and convolutional neural network [J]. Telecommunications Science, 2021, 37(3): 133-145.
[3]	Jingcheng CAO, Jidong ZHANG, Guojie SHI. A method for improving recall rate of similar image retrieval by using edge enhancement technology [J]. Telecommunications Science, 2021, 37(1): 76-84.
[4]	Qiang LI,Zilu KANG. Video crowd counting method based on conv-pooling deep spatial and temporal features [J]. Telecommunications Science, 2018, 34(6): 72-79.
[5]	Huijun XING,Shuo CHANG. Pedestrian surveillance system based on mobile vehicle [J]. Telecommunications Science, 2017, 33(2): 120-127.