Integration algorithm of improved maximum a posteriori probability vector quantization and least squares support vector machine

doi:10.11772/j.issn.1001-9081.2015.07.2101

Abstract

Abstract:

In view of the current efficiency problem of speaker recognition system, this paper utilized the tactics of integration algorithm to put forward a new kind of speaker recognition system framework. The traditional Maximum A posteriori Probability Vector Quantization (VQ-MAP) algorithm only focuses on the average vector regardless of weight. In order to solve this problem, this paper put forward an improved algorithm based on VQ-MAP. The algorithm used weighted average vector instead of average vector. Moreover, Support Vector Machine (SVM) algorithm costs too much time, so Least Squares Support Vector Machine (LS-SVM) was used instead of SVM. Finally, in the speaker recognition system, this paper used the parameters calculated from the improved VQ-MAP algorithm as training set of LS-SVM. The experimental results show that, the modeling time of integration algorithm based on improved VQ-MAP and LS-SVM is about 40% less than that of traditional SVM algorithm when using the Radial Basis Function (RBF) kernel function and the sample of 40 people. As the threshold value is 1 and the test speech time is 4 s, compared to the traditional VQ-MAP and SVM algorithm, the deterrent rate is reduced by 1.1%, the false rejection rate is reduced by 2.9% and the recognition rate is increased by 3.9%. As the threshold value is 1 and the test speech time is 4 s, compared to the traditional VQ-MAP and LS-SVM algorithm, the deterrent rate is reduced by 3.6%, the false rejection rate is reduced by 2.7% and the recognition rate is increased by 4.4%. The results show that the integrated algorithm can improve the recognition rate effectively and reduce the operation time significantly, meanwhile reduce the deterrent rate and the false rejection rate.

Key words: Maximum A posteriori Probability (MAP), Least Squares Support Vector Machine (LS-SVM), weight, mean vector, speaker recognition

摘要：

针对目前说话人识别系统的效率问题,采用集成算法的策略,提出一种新的说话人识别系统框架。首先,考虑到传统的最大后验概率矢量量化(VQ-MAP)算法中只关注平均矢量而不考虑权重的问题,提出了改进的VQ-MAP算法,使用加权平均向量来代替平均向量;然后,由于支持向量机(SVM)算法相对耗时,故采用最小二乘支持向量机(LS-SVM)替代SVM算法;最后,在说话人识别系统中,利用改进的VQ-MAP算法所得参数集作为LS-SVM的训练样本。实验结果表明,基于改进的VQ-MAP和LS-SVM的集成算法,与传统的SVM算法相比,在均使用径向基函数(RBF)核函数时,对40人样本数据建模时间上减少接近40%;在阈值为1,测试语音时长为4 s时,与传统的VQ-MAP和SVM算法相比,误识率降低了1.1%,误拒率降低了2.9%,识别率提高了3.9%;在阈值为1,测试语音时长为4 s时,与传统的VQ-MAP和LS-SVM算法相比,误识率降低了3.6%,误拒率降低了2.7%,识别率提高了4.4%。结果表明,集成算法能够有效提高算法识别率,明显减少运算时间,同时降低误识率和误拒率。

关键词: 最大后验概率, 最小二乘支持向量机, 权重, 平均向量, 说话人识别

CLC Number:

ZHANG Jun, GUAN Shengxiao. Integration algorithm of improved maximum a posteriori probability vector quantization and least squares support vector machine[J]. Journal of Computer Applications, 2015, 35(7): 2101-2104.

张俊, 关胜晓. 基于改进的最大后验概率矢量量化和最小二乘支持向量机集成算法[J]. 计算机应用, 2015, 35(7): 2101-2104.

References

[1] HAN J, ZHANG L, ZHENG T. Speech signal processing [M]. Beijing: Tsinghua University Press, 2004: 191-193.(韩纪庆,张磊,郑铁然.语音信号处理[M].北京:清华大学出版社,2004:191-193.)
[2] GEORGE K K, ARUNRAJ K, SREEKUMAR K T, et al. Towards improving the performance of text/language independent speaker recognition systems [C]// EPSCICON 2014: Proceedings of the 2014 International Conference on Power Signals Control and Computations. Piscataway: IEEE, 2014: 1-6.
[3] MENG J, YANG D. Research on training duration of UBM in speaker identification [J]. Journal of Beijing Information Science and Technology University, 2013, 28(3): 87-91.(孟君,杨大利.说话人辨认中通用背景模型训练时长研究[J].北京信息科技大学学报:自然科学版,2013,28(3):87-91.)
[4] ZHOU Y. Speaker verification based on factor analysis [D].Hefei: University of Science and Technology of China, 2010.(周毓.基于因子分析的说话人确认[D].合肥:中国科学技术大学,2010.)
[5] HAUTAMAKI V, KINNUNEN T, KARKKAINEN I, et al. Maxi-mum a posteriori adaptation of the centroid model for speaker verification [J]. IEEE Signal Processing Letters, 2008, 15: 162-165.
[6] ZHAN L, JING X. Speaker recognition system based on VQ-MAP and SVM [J]. Computer Engineering and Applications, 2011, 47(13): 136-138.(展领,景新幸.基于VQ-MAP和SVM融合的说话人识别系统[J].计算机工程与应用,2011,47(13):136-138.)
[7] SUYKENS J A K, VANDEWALLE J. Least squares support vector machine classifiers [J]. Neural Processing Letters, 1999, 9(3): 293-300.
[8] ZHENG R, ZHANG S, XU B. Text-independent speaker identification using GMM-UBM and frame level likelihood normalization [C]// Proceedings of the 2004 International Symposium on Chinese Spoken Language Processing. Piscataway: IEEE, 2004: 289-292.
[9] SHA F, SAUL L K. Large margin Gaussian mixture modeling for phonetic classification and recognition [C]// ICASSP 2006: Proceedings of the 2006 IEEE International Conference on Acoustics, Speech and Signal Processing. Piscataway: IEEE, 2006: 265-268.
[10] KINNUNEN T, FRÄNTI P. Speaker discriminative weighting method for VQ-based speaker identification [C]// Proceedings of the Third International Conference on Audio- and Video-based Biometric Person Authentication, LNCS 2091. Berlin: Springer, 2001: 150-156.
[11] DAN Z, ZHENG S, SUN S, et al. Speaker recognition based on LS-SVM [C]// ICICIC'08: Proceedings of the 3rd International Conference on Innovative Computing Information and Control. Piscataway: IEEE, 2008: 525-525.
[12] KANUNGO T, MOUNT D M, NETANYAHU N S, et al. An efficient k-means clustering algorithm: analysis and implementation [J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2002, 24(7): 881-892.
[13] MURTY K S R, YEGNANARAYANA B. Combining evidence from residual phase and MFCC features for speaker recognition [J]. IEEE Signal Processing Letters, 2006, 13(1): 52-55.
[14] HATCH A O, KAJAREKAR S S, STOLCKE A. Within-class covariance normalization for SVM-based speaker recognition [EB/OL]. [2015-01-10]. http://www.icsi.berkeley.edu/pubs/speech/HatchICSLP06.pdf.
[15] YANG J, ZHANG L, YANG J, et al. From classifiers to discriminators: a nearest neighbor rule induced discriminant analysis [J]. Pattern Recognition, 2011, 44(7): 1387-1402.
[16] HANSEN J H L, SUH J W, ANGKITITRAKUL P, et al. Effective background data selection for SVM-based speaker recognition with unseen test environments: more is not always better [J]. International Journal of Speech Technology, 2014, 17(3): 211-221.
[17] SUYKENS J A K, de BRABANTER J, LUKAS L, et al. Weighted least squares support vector machines: robustness and sparse approximation [J]. Neurocomputing, 2002, 48(1): 85-105.
[18] GAO H, WANG X. LS-SVM based intrusion detection using kernel space approximation and kernel-target alignment [C]// WCICA 2006: Proceedings of the Sixth World Congress on Intelligent Control and Automation. Piscataway: IEEE, 2006: 4214-4218.
[19] CAICEDO A, van HUFFEL S. Weighted LS-SVM for function estimation applied to artifact removal in bio-signal processing [C]// EMBC 2010: Proceedings of the 2010 Annual International Conference on Engineering in Medicine and Biology Society. Piscataway: IEEE, 2010: 988-991.