Forged Voice Identification Method Based on Feature Fusion and Multi-channel GRU

doi:10.3969/j.issn.1671-1122.2021.10.001

Abstract

Abstract:

In order to solve the problems of poor generalization ability and low detection accuracy of existing counterfeit authentication models, this article proposes a three-channel GRU forged voice identification model based on hybrid feature fusion. Validated on the ASVspoof2019 dataset, the accuracy of the proposed method reaches 96.30% for the detection of fake Logical Access samples and 87.33% for that of the fake Physical Access samples, which is better than other algorithms. The experimental results prove that the fake voice detection method based on time-frequency domain feature fusion can learn more effective authenticity identification features and obtain higher detection accuracy.

Key words: speech forgery detection, multi-channel GRU, feature fusion, deep learning

CLC Number:

TP309

PAN Xiaoqin, DU Yanhui. Forged Voice Identification Method Based on Feature Fusion and Multi-channel GRU[J]. Netinfo Security, 2021, 21(10): 1-7.

Figures/Tables 8

References 16

[1]	HANILI C, KINNUNEN T, SAHIDULLAH M, et al. Classifiers for Synthetic Speech Detection: A Comparison [C]//ISCA. Interspeech 2015, September 6-10, 2015, Dresden, Germany. Dresden: ISCA, 2015: 2057-2061.
[2]	SAHIDULLAH M, KINNUNEN T, HANILCI C. A Comparison of Features for Synthetic Speech Detection [C]//ISCA. Interspeech 2015, September 6-10, 2015, Dresden, Germany. Dresden: ISCA, 2015: 2087-2091.
[3]	CHEN Lianwu, GUO Wu, DAI Lirong. Speaker Verification Against Synthetic Speech [C]//IEEE. 7th International Symposium on Chinese Spoken Language Processing, November 29-December 3, 2010, Tainan, Taiwan, China. New Jersey: IEEE, 2010: 309-312.
[4]	YANG Jichen, HE Qianhua, HU Yongjian, et al. CBC-based Synthetic Speech Detection[J]. International Journal of Digital Crime and Forensics (IJDCF), 2019, 11(2):63-74.
[5]	TIAN Xiaohai, DU S, XIAO Xiong, et al. Detecting Synthetic Speech Using Long Term Magnitude and Phase Information [C]//IEEE. China Summit and International Conference on Signal and Information Processing (ChinaSIP), July 12-15, 2015, Chengdu, China. Chengdu: IEEE, 2015: 611-615.
[6]	WU Zhizheng, XIAO Xiong, CHNG E S, et al. Synthetic Speech Detection Using Temporal Modulation Feature [C]//IEEE. International Conference on Acoustics, Speech and Signal Processing, May 26-31, 2013, Vancouver, BC, Canada. Vancouver: IEEE, 2013: 7234-7238.
[7]	SAILOR H B, KAMBLE M R, PATIL H A. Unsupervised Representation Learning Using Convolutional Restricted Boltzmann Machine for Spoof Speech Detection [C]//ISCA. Interspeech 2017, August 20-24, 2017, Stockholm, Sweden. Stockholm: ISCA, 2017: 2601-2605.
[8]	TIAN Xiaohai, XIAO Xiong, CHNG E S, et al. Spoofing Speech Detection Using Temporal Convolutional Neural Network [C]//IEEE. 2016 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, December 13-16, 2016, Jeju, Korea(South). Jeju: IEEE, 2016: 1-6.
[9]	WANG Run, JUEFEI-XU F, HUANG Yihao, et al. DeepSonar: Towards Effective and Robust Detection of AI-synthesized Fake Voices [C]//ACM. 28th ACM International Conference on Multimedia, October 12-16, 2020, Seattle, WA, USA. New York: ACM, 2020: 1207-1216.
[10]	WU Zhenzong, DAS R K, YANG Jichen, et al. Light Convolutional Neural Network with Feature Genuinization for Detection of Synthetic Speech Attacks [C]//ISCA. Interspeech 2020, October 25-19, 2020, Shanghai, China. Shanghai: ISCA, 2020: 1101-1105.
[11]	LI Xu, LI Na, WENG Chao, et al. Replay and Synthetic Speech Detection with Res2Net Architecture [C]//IEEE. International Conference on Acoustics, Speech and Signal Processing (ICASSP), June 6-11, 2021, Toronto, Ontario, Canada. Toronto: IEEE, 2021: 6354-6358.
[12]	GOMEZ-ALANIS A, PEINADO A M, GONZALEZ J A, et al. A Deep Identity Representation for Noise Robust Spoofing Detection [C]//ISCA. Interspeech 2018, September 2-6, 2018, Hyderabad, India. Hyderabad: ISCA, 2018: 676-680.
[13]	DUA M, JAIN C, KUMAR S. LSTM and CNN Based Ensemble Approach for Spoof Detection Task in Automatic Speaker Verification Systems[J]. Journal of Ambient Intelligence and Humanized Computing, 2021, 12(2):1-16. doi: 10.1007/s12652-020-02846-7 URL
[14]	GOMEZ-ALANIS A, PEINADO A M, GONZALEZ J A, et al. A Light Convolutional GRU-RNN Deep Feature Extractor for ASV Spoofing Detection [C]//ISCA. Interspeech 2019, September 15-19, 2019, Graz, Austria. Graz: ISCA, 2019: 1068-1072.
[15]	KAMBLE M R, PATIL H A. Detection of Replay Spoof Speech Using Teager Energy Feature Cues[EB/OL]. https://www.researchgate.net/publication/343659925_Detection_of_Replay_Spoof_Speech_Using_Teager_Energy_Feature_Cues, 2021-05-01.
[16]	VARIANI E, SAINATH T N, SHAFRAN I, et al. Complex Linear Projection (CLP): A Discriminative Approach to Joint Feature Extraction and Acoustic Modeling [C]//ISCA. Interspeech 2016, September 8-12, 2016, San Francisco, CA, USA. San Francisco: ISCA, 2016: 808-812.

数据集分类	物理方式		逻辑方式
数据集分类	真实样本数	伪造样本数	真实样本数	伪造样本数
训练集	2500	2500	2500	2500
验证集	500	500	500	500
测试集	1500	1500	1500	1500

名称	信息
CPU	Intel(R) Core(TM) i9-10920X CPU @3.50 GHz
GPU	NVIDIA GeForce RTX 3090
内存	32 GB
操作系统	Ubuntu 16.04.6 LTS
Python版本	3.6.13
深度学习库	Tensorflow 2.1.0, keras 2.3.1

模型	ACC	EER	AUC	F1-score	Precision	Recall
RNN	88.87	11.98	84.64	89.84	83.64	97.04
LSTM	89.27	11.91	91.09	89.65	87.73	91.66
GRU	88.93	7.17	97.65	89.92	83.58	97.31
Multi-channel RNN	92.70	7.10	98.13	92.77	93.29	92.25
Multi-channel LSTM	94.27	5.82	93.20	94.37	94.00	94.74
Multi-channel GRU	96.30	0.34	99.99	96.48	93.20	100.00

模型	ACC	EER	AUC	F1-score	Precision	Recall
RNN	79.87	21.44	96.45	78.84	83.83	74.40
LSTM	82.43	19.35	92.99	81.52	86.78	76.85
GRU	81.70	18.41	94.11	80.98	85.02	77.31
Multi-channel RNN	83.77	15.26	88.75	82.13	92.25	74.01
Multi-channel LSTM	83.99	14.07	93.22	84.95	80.88	89.46
Multi-channel GRU	87.33	7.92	97.18	88.48	82.11	95.93