基因表达式编程与HMM融合技术应用研究
|
摘要
随着当前智能技术的飞速发展,各种智能方法在机器学习,自然语言理解,语音识别,数据挖掘,自动控制等各个领域得到广泛使用。隐马尔可夫模型作为一种统计模型,在人机交互,机器学习,语音识别,文本信息抽取等领域中获得了重大突破。
隐马尔可夫模型(Hidden Markov Model,HMM)是一种具有学习能力的统计模型,在语音与信息处理诸多领域已经得到了成功的应用,其HMM的参数设置是其能否成功应用的关键,故提取特征矢量和确定HMM状态输出概率密度函数形式是参数设置的重要环节,传统HMM的Baum-Welch算法本质上是一个梯度下降算法,该算法的优点就是收敛速度快,但是在参数估计时极易陷入局部最优从而影响最终的效果。
而作为演化算法家族的新成员的基因表达式编程是在遗传算法和遗传编程的基础上被提出来的,它结合了遗传算法和遗传编程设计的优点,并克服了它们的不足,是遗传算法和遗传编程的继承和发展,在函数发现,参数优化,数学建模等方面较其他进化算法有着明显优势,目前已经成为国际进化计算领域的研究热点。
基于以上原因,考虑将HMM和GEP两者的优点融合在一起进行深入研究,将最新的研究应用到语音识别算法,文本信息抽取,建模等领域以期取得更好的效果。
本文对HMM的训练算法和GEP建模方法进行了研究,主要研究成果如下:
(1)对HMM的基本训练算法进行了理论分析,并给出了形式化定义,并在这些定义的基础上,就HMM的三种算法进行具体分析,分析了常用的HMM算法的优点和缺点,并提出自己的见解。
(2)在Candida Ferreira方法的基础上对GEP-PO算法用来优化参数的染色体构造给出了形式化定义,系统分析了如何用该算法进行参数的优化,并对该方法的优点和缺点进行了分析。提出并实现了一种高效的GEP优化HMM的训练算法(GEP-PO-BW-based HMM Algorithms ,GBHA)。
(3)对传统演化算法存在用网络反馈信息导致学习速度较慢的不足,提出一种基于HMM的GEP建模算法(Time Series Prediction with GEP Based on HMM,H-GEPTSP)。
(4)通过实验验证了上述各算法的高效性并取得的预期效果。
With the current rapid development of smart technology, all kinds of intelligent methods in machine learning, natural language understanding, automatic control and other fields have been widely used. Hidden Markov Model, as a statistical model, in human-computer interaction, machine learning, speech recognition and other areas was a major breakthrough.
Evolutionary algorithm as a new family of gene expression programming in genetic algorithms and genetic programming was put forward basis, which combines genetic algorithms and genetic programming design advantages, and overcome their deficiencies, genetic algorithms and genetic continuation and development of programming, was found in the function, parameter optimization, mathematical modeling and other methods than the other evolutionary algorithm has obvious advantages, has become an international research focus in the field of evolutionary computation. HMM (Hidden Markov Model, HMM) is a statistical model with learning ability, in many areas of voice and information processing has been applied successfully, the HMM parameter set is the key to the success of applications, Therefore, vector feature extraction and identification of HMM state output probability density function of the form is an important part of parameter settings, the traditional HMM-Baum-Welch algorithm is essentially a gradient descent algorithm is the advantage of faster convergence, but the parameter estimates very easy to fall into local optimum and thus affect the final results. For these reasons, consider the advantages of both HMM and GEP combines in-depth study of the latest speech recognition algorithm applied to the areas of modeling in order to achieve better results. Global search of Gene Expression Programming (Gene Expression Programming, GEP) A key feature is the ability to efficiently find the global optimal solution quickly, this paper, the GEP is introduced into the training of HMM to put forward an improved training methods (GEP -PO-BW-based HMM Algorithms, GBHA), finally realized by Matlab simulation under the above-mentioned algorithm, finally got a better method to improve the system efficiency and stability. In this paper, HMM training algorithms and modeling methods of GEP, the main research results are as follows:
(1) of the HMM training algorithm is the basic theory and gives a formal definition and the basis of these definitions, the three algorithms on HMM specific analysis, and with the dynamic time warping (DWT) technology, artificial neural network (Artificial Neural Network, ANN) were compared, analysis of common strengths and weaknesses of HMM algorithm, and made their views.
(2) Candida Ferreira method based on the GEP-PO algorithm used to optimize the parameters of the chromosome structure is given a formal definition, systems analysis of how to optimize the algorithm parameters, and the advantages and disadvantages of this method were analyzed. Proposed and implemented an efficient optimization of HMM training algorithm GEP (GEP-PO-BW-based HMM Algorithms, GBHA).
(4) the existence of the evolutionary algorithm in learning with a network of feedback led to less than slower, the GEP proposes a HMM-based modeling algorithm (Time Series Prediction with GEP Based on HMM, H-GEPTSP).
(5) integration of the experiment verify the efficiency of various algorithms and achieved the expected results.
隐马尔可夫模型(Hidden Markov Model,HMM)是一种具有学习能力的统计模型,在语音与信息处理诸多领域已经得到了成功的应用,其HMM的参数设置是其能否成功应用的关键,故提取特征矢量和确定HMM状态输出概率密度函数形式是参数设置的重要环节,传统HMM的Baum-Welch算法本质上是一个梯度下降算法,该算法的优点就是收敛速度快,但是在参数估计时极易陷入局部最优从而影响最终的效果。
而作为演化算法家族的新成员的基因表达式编程是在遗传算法和遗传编程的基础上被提出来的,它结合了遗传算法和遗传编程设计的优点,并克服了它们的不足,是遗传算法和遗传编程的继承和发展,在函数发现,参数优化,数学建模等方面较其他进化算法有着明显优势,目前已经成为国际进化计算领域的研究热点。
基于以上原因,考虑将HMM和GEP两者的优点融合在一起进行深入研究,将最新的研究应用到语音识别算法,文本信息抽取,建模等领域以期取得更好的效果。
本文对HMM的训练算法和GEP建模方法进行了研究,主要研究成果如下:
(1)对HMM的基本训练算法进行了理论分析,并给出了形式化定义,并在这些定义的基础上,就HMM的三种算法进行具体分析,分析了常用的HMM算法的优点和缺点,并提出自己的见解。
(2)在Candida Ferreira方法的基础上对GEP-PO算法用来优化参数的染色体构造给出了形式化定义,系统分析了如何用该算法进行参数的优化,并对该方法的优点和缺点进行了分析。提出并实现了一种高效的GEP优化HMM的训练算法(GEP-PO-BW-based HMM Algorithms ,GBHA)。
(3)对传统演化算法存在用网络反馈信息导致学习速度较慢的不足,提出一种基于HMM的GEP建模算法(Time Series Prediction with GEP Based on HMM,H-GEPTSP)。
(4)通过实验验证了上述各算法的高效性并取得的预期效果。
With the current rapid development of smart technology, all kinds of intelligent methods in machine learning, natural language understanding, automatic control and other fields have been widely used. Hidden Markov Model, as a statistical model, in human-computer interaction, machine learning, speech recognition and other areas was a major breakthrough.
Evolutionary algorithm as a new family of gene expression programming in genetic algorithms and genetic programming was put forward basis, which combines genetic algorithms and genetic programming design advantages, and overcome their deficiencies, genetic algorithms and genetic continuation and development of programming, was found in the function, parameter optimization, mathematical modeling and other methods than the other evolutionary algorithm has obvious advantages, has become an international research focus in the field of evolutionary computation. HMM (Hidden Markov Model, HMM) is a statistical model with learning ability, in many areas of voice and information processing has been applied successfully, the HMM parameter set is the key to the success of applications, Therefore, vector feature extraction and identification of HMM state output probability density function of the form is an important part of parameter settings, the traditional HMM-Baum-Welch algorithm is essentially a gradient descent algorithm is the advantage of faster convergence, but the parameter estimates very easy to fall into local optimum and thus affect the final results. For these reasons, consider the advantages of both HMM and GEP combines in-depth study of the latest speech recognition algorithm applied to the areas of modeling in order to achieve better results. Global search of Gene Expression Programming (Gene Expression Programming, GEP) A key feature is the ability to efficiently find the global optimal solution quickly, this paper, the GEP is introduced into the training of HMM to put forward an improved training methods (GEP -PO-BW-based HMM Algorithms, GBHA), finally realized by Matlab simulation under the above-mentioned algorithm, finally got a better method to improve the system efficiency and stability. In this paper, HMM training algorithms and modeling methods of GEP, the main research results are as follows:
(1) of the HMM training algorithm is the basic theory and gives a formal definition and the basis of these definitions, the three algorithms on HMM specific analysis, and with the dynamic time warping (DWT) technology, artificial neural network (Artificial Neural Network, ANN) were compared, analysis of common strengths and weaknesses of HMM algorithm, and made their views.
(2) Candida Ferreira method based on the GEP-PO algorithm used to optimize the parameters of the chromosome structure is given a formal definition, systems analysis of how to optimize the algorithm parameters, and the advantages and disadvantages of this method were analyzed. Proposed and implemented an efficient optimization of HMM training algorithm GEP (GEP-PO-BW-based HMM Algorithms, GBHA).
(4) the existence of the evolutionary algorithm in learning with a network of feedback led to less than slower, the GEP proposes a HMM-based modeling algorithm (Time Series Prediction with GEP Based on HMM, H-GEPTSP).
(5) integration of the experiment verify the efficiency of various algorithms and achieved the expected results.
引文
[1]倪崇嘉,刘文举,徐波,汉语大词汇量连续语音识别系统研究进展,中文信息学报[J]2009/23(01)
[2]雷建军,杨震,噪声鲁棒语音识别研究综述,计算机应用研究[J], 2009年26(04)1210-1216
[3]李津涛.语音特征参数提取的仿真研究中国新通信[J].2009年第9期52-54
[4] Ferreira C. Gene Expression Programming: Mathematical Modeling by an Artificial Intelligence [M],Second ed.,Springer-Verlag, 2006.
[5] EPHRAIM Y, COHEN I. Recent advancements in speech enhancement [ K ]//The electrical engineering handbook. [ S. l. ] : CRC Press, 2006.
[6] Nikola Trbovica, Felician Datlceaa, Thomas Langerb, Using wavelet de -noised spectra in NMR screening[J]. Journal of Magnetic Resonance 2005, 173 (2) : 280 - 287.
[7]卢坚,基于隐马尔可夫模型的音频自动分类,软件学报[J] 2002/13 (08)1593-05
[8]曾剑平,郭东辉.一种基于HMM和遗传算法的伪装入侵检测方法.小型微型计算机系统[J] 2007, 28(07):1210~1215
[9]王仁华.中文语音交互技术标准化工作进展[J].信息技术标准化,2004,(3):4-5.
[10] oza, John, R. Genetic Programming, On the Programming of Computers by Means of Natural Selection. 1st Ed. MIT Press 1992.
[11]徐哲,白焰.遗传编程[J],自动化仪表,2002,23(10):1-7.
[12] Yorick Hardy.Gene Expression Programming and One-Dimensional Chaotic Maps[J],,International Journal of Modern Physics C,2002 Vol.13(1):13-24
[13] Zhou Chi,Xiao Weimin,Tirpak Thomas M.,et al.Evolving accurate and compact classification rules with gene expression programming[J],.IEEE Transactions on Evolutionary Computation,2003,7(6):519-531.
[14] Comparative analysis of using artificial neural networks (ANN) and gene expression programming (GEP) in backcalculation of pavement layer thickness[J], Indian Journal of Engineering & Materials Sciences Vol. 12, February 2005, pp
[15] Xin Li, Chi Zhou, Weimin Xiao,Peter C. Nelson. Prefix Gene Expression Programming[C], Genetic and Evolutionary Computation Conference (GECCO)’05, June 25-29, 2005,Washington, DC, USA.
[16] Ozlem Terzi and M.Erol Keskin. Evaporation Estimation Using Gene Expression Programming[J]. Journal of Applied Sciences 5 (3): 508-512, 2005
[17] Robert Gempeler, Image Compression Using Gene Expression Programming. Technical Report, NSF Research Experiences for Undergraduates, Computer Vision and Image Processing, Department of Computer Science, Utah State University, 2006
[18] Stewart W. Wilson .Classifier Conditions Using Gene Expression Programming[J], Lecture Notes In Artificial Intelligence,2008.pp 206 - 217
[19] Jose G.Moreno-Torres,Xavier Llor_a and David E.Goldberg. Binary Representation in Gene ExpressionProgramming: Towards a Better Scalability ,IlliGAL Technical Report No. 2008003,February, 2009.
[20] Zuo Jie, Tang Changjie, Zhang Tianqing. Mining Predicate Association Rule by Gene Expression Programming [A], Meng Xiaofeng, Su Jianwne, Wang Yujun. LNCS (Lecture Notes In Computer science)[C]. WAIM02 (International Conference for Web Information Age 2002), Springer Verlag Berling Heidelberg 2002, 2419: 92-103.
[21] Zuo Jie, Tang Chang-jie, Li Chuan, etc. Time Series Prediction based on Gene Expression Programming, WAIM04 (International Conference for Web Information Age 2004). LNCS (Lecture Notes In Computer science) Vol.3129, 55-64, edited by Q Li and G. Wang, Springer Verlag Berling Heidelberg, 2004, 8.
[22]段磊,唐常杰,左劼，等.基于基因表达式编程的抗噪声数据的函数发现方法[J],计算机研究与发展,41(10), 2004, 1684-1689.
[23]元昌安,唐常杰,左劼，等.基于基因表达式编程的函数发现-收敛性分析与残差制导进化算法[J],四川大学学报(工程科学版),Vol.36 No.6, 2004, 100-105.
[24]元昌安,唐常杰,温远光,等.基于基因表达式编程的智能模型库系统的实现[J].四川大学学报(工程科学版), Vol.37 No.3, 2005, 99-104.
[25]蒋思伟,蔡之华,曾丹,等.基于模拟退火的并行基因表达式编程算法研究[J].电子学报,Vol.33 No.11 2005, 2017-2021.
[26]钟义啸,唐常杰,陈宇,等.提高基因表达式编程发现知识效率的回溯策略[J].四川大学学报(自然科学版),Vol.43 No.2 2006, 299-304.
[27]张欢.基因表达式编程中的转基因关键技术研究[D].成都:四川大学, 2006.
[28]邓松.基于GEP的空间数据函数发现及其在分类中的应用,广西师范学院硕士学位论文,2006.
[29]司宏宗.基因表达式编程与支持向量机在疾病诊断和QSAR/QSPR中的应用研究[D],兰州:兰州大学,2006.
[30]彭京,唐常杰,元昌安,朱明放,乔少杰.基于重叠表达的多基因进化算法[J],计算机学报,2007,30(5):776-785.
[31]向勇,唐常杰等,基于基因表达式编程的多目标优化算法[J],四川大学学报(工程科学版), 2007,39(4):124-129
[32]颜雪松,时晨,黄士坦.基于基因表达式程序设计的电路优化算法研究[J],微电子学与计算机,2008,25(1):120-122
[33]颜雪松,时晨,黄士坦.基于基因表达式程序设计的电路优化算法研究[J],微电子学与计算机,2008,25(1):120-122
[34] Jia LV . Study on Chaos Immune Network Algorithm for Multimodal Function Optimization[C],Proceedings of the Fourth International Conference on Fuzzy Systems and Knowledge Discovery(FSKD 2007),2007:1-5
[35]黄隆胜,廖颀. GEP软件设计及其K表达式快速求值算法[J].计算机工程与设计,2007,28(4):775-776.
[36] G. Harik,F. G. Lobo,and D.E.Goldberg.The compact genetic algorithm[J], IEEE Trans. Evol Comput ,1999,3(4):287–297.
[37]张长胜,HMM在语音识别中的应用研究,吉林大学硕士学位论文,2006
[38]杨凤芹,基于粒子群的优化方法研究,2009
[39]马龙华,车载环境下语音识别方法研究,哈尔滨工程大学博士学位论文,2008
[40]刘畅,改进的HMM训练方法在语音识别中的应用,吉林大学硕士学位论文,2007.
[41]丁琼,网络化智能控制系统研究,贵州大学硕士学位论文,2008
[42]赵洪华,输油管线破坏预防监测及其定位技术的研究,济南大学硕士学位论文,2006
[43]张耀兰,语音识别技术在导航设备中的应用,北京交通大学硕士学位论文,2009
[44]张志霞,语音识别中个人特征参数提取研究,中北大学硕士学位论文,2009
[45]王一平用遗传算法改进HMM的语音识别算法研究太原理工大学硕士学位论文,2007
[46]刘涛,管道检测系统中的信号处理技术研究,天津大学硕士学位论文,2007
[47]徐毅琼,人脸识别技术研究,中国人民解放军信息工程大学硕士学位论文,2005
[48]张志刚,基于神经网络/HMM的语音识别算法的研究,武汉理工大学硕士学位论文,2006
[49]李建宁,汉语孤立词语音识别的研究与实现,西北大学硕士学位论文,2007
[50]项静恬等,多种时序模型的建模比较,数理统计与管理,1991年2月
[51] George,G.S.Forecasting chaotic time series with genetic algorithms. Physical Review E,1997,55(3)
[52]袁俊. HMM连续语音识别中Viterbi算法的优化及应用.电子技术[J]2001年第28卷第期48-51。
[2]雷建军,杨震,噪声鲁棒语音识别研究综述,计算机应用研究[J], 2009年26(04)1210-1216
[3]李津涛.语音特征参数提取的仿真研究中国新通信[J].2009年第9期52-54
[4] Ferreira C. Gene Expression Programming: Mathematical Modeling by an Artificial Intelligence [M],Second ed.,Springer-Verlag, 2006.
[5] EPHRAIM Y, COHEN I. Recent advancements in speech enhancement [ K ]//The electrical engineering handbook. [ S. l. ] : CRC Press, 2006.
[6] Nikola Trbovica, Felician Datlceaa, Thomas Langerb, Using wavelet de -noised spectra in NMR screening[J]. Journal of Magnetic Resonance 2005, 173 (2) : 280 - 287.
[7]卢坚,基于隐马尔可夫模型的音频自动分类,软件学报[J] 2002/13 (08)1593-05
[8]曾剑平,郭东辉.一种基于HMM和遗传算法的伪装入侵检测方法.小型微型计算机系统[J] 2007, 28(07):1210~1215
[9]王仁华.中文语音交互技术标准化工作进展[J].信息技术标准化,2004,(3):4-5.
[10] oza, John, R. Genetic Programming, On the Programming of Computers by Means of Natural Selection. 1st Ed. MIT Press 1992.
[11]徐哲,白焰.遗传编程[J],自动化仪表,2002,23(10):1-7.
[12] Yorick Hardy.Gene Expression Programming and One-Dimensional Chaotic Maps[J],,International Journal of Modern Physics C,2002 Vol.13(1):13-24
[13] Zhou Chi,Xiao Weimin,Tirpak Thomas M.,et al.Evolving accurate and compact classification rules with gene expression programming[J],.IEEE Transactions on Evolutionary Computation,2003,7(6):519-531.
[14] Comparative analysis of using artificial neural networks (ANN) and gene expression programming (GEP) in backcalculation of pavement layer thickness[J], Indian Journal of Engineering & Materials Sciences Vol. 12, February 2005, pp
[15] Xin Li, Chi Zhou, Weimin Xiao,Peter C. Nelson. Prefix Gene Expression Programming[C], Genetic and Evolutionary Computation Conference (GECCO)’05, June 25-29, 2005,Washington, DC, USA.
[16] Ozlem Terzi and M.Erol Keskin. Evaporation Estimation Using Gene Expression Programming[J]. Journal of Applied Sciences 5 (3): 508-512, 2005
[17] Robert Gempeler, Image Compression Using Gene Expression Programming. Technical Report, NSF Research Experiences for Undergraduates, Computer Vision and Image Processing, Department of Computer Science, Utah State University, 2006
[18] Stewart W. Wilson .Classifier Conditions Using Gene Expression Programming[J], Lecture Notes In Artificial Intelligence,2008.pp 206 - 217
[19] Jose G.Moreno-Torres,Xavier Llor_a and David E.Goldberg. Binary Representation in Gene ExpressionProgramming: Towards a Better Scalability ,IlliGAL Technical Report No. 2008003,February, 2009.
[20] Zuo Jie, Tang Changjie, Zhang Tianqing. Mining Predicate Association Rule by Gene Expression Programming [A], Meng Xiaofeng, Su Jianwne, Wang Yujun. LNCS (Lecture Notes In Computer science)[C]. WAIM02 (International Conference for Web Information Age 2002), Springer Verlag Berling Heidelberg 2002, 2419: 92-103.
[21] Zuo Jie, Tang Chang-jie, Li Chuan, etc. Time Series Prediction based on Gene Expression Programming, WAIM04 (International Conference for Web Information Age 2004). LNCS (Lecture Notes In Computer science) Vol.3129, 55-64, edited by Q Li and G. Wang, Springer Verlag Berling Heidelberg, 2004, 8.
[22]段磊,唐常杰,左劼，等.基于基因表达式编程的抗噪声数据的函数发现方法[J],计算机研究与发展,41(10), 2004, 1684-1689.
[23]元昌安,唐常杰,左劼，等.基于基因表达式编程的函数发现-收敛性分析与残差制导进化算法[J],四川大学学报(工程科学版),Vol.36 No.6, 2004, 100-105.
[24]元昌安,唐常杰,温远光,等.基于基因表达式编程的智能模型库系统的实现[J].四川大学学报(工程科学版), Vol.37 No.3, 2005, 99-104.
[25]蒋思伟,蔡之华,曾丹,等.基于模拟退火的并行基因表达式编程算法研究[J].电子学报,Vol.33 No.11 2005, 2017-2021.
[26]钟义啸,唐常杰,陈宇,等.提高基因表达式编程发现知识效率的回溯策略[J].四川大学学报(自然科学版),Vol.43 No.2 2006, 299-304.
[27]张欢.基因表达式编程中的转基因关键技术研究[D].成都:四川大学, 2006.
[28]邓松.基于GEP的空间数据函数发现及其在分类中的应用,广西师范学院硕士学位论文,2006.
[29]司宏宗.基因表达式编程与支持向量机在疾病诊断和QSAR/QSPR中的应用研究[D],兰州:兰州大学,2006.
[30]彭京,唐常杰,元昌安,朱明放,乔少杰.基于重叠表达的多基因进化算法[J],计算机学报,2007,30(5):776-785.
[31]向勇,唐常杰等,基于基因表达式编程的多目标优化算法[J],四川大学学报(工程科学版), 2007,39(4):124-129
[32]颜雪松,时晨,黄士坦.基于基因表达式程序设计的电路优化算法研究[J],微电子学与计算机,2008,25(1):120-122
[33]颜雪松,时晨,黄士坦.基于基因表达式程序设计的电路优化算法研究[J],微电子学与计算机,2008,25(1):120-122
[34] Jia LV . Study on Chaos Immune Network Algorithm for Multimodal Function Optimization[C],Proceedings of the Fourth International Conference on Fuzzy Systems and Knowledge Discovery(FSKD 2007),2007:1-5
[35]黄隆胜,廖颀. GEP软件设计及其K表达式快速求值算法[J].计算机工程与设计,2007,28(4):775-776.
[36] G. Harik,F. G. Lobo,and D.E.Goldberg.The compact genetic algorithm[J], IEEE Trans. Evol Comput ,1999,3(4):287–297.
[37]张长胜,HMM在语音识别中的应用研究,吉林大学硕士学位论文,2006
[38]杨凤芹,基于粒子群的优化方法研究,2009
[39]马龙华,车载环境下语音识别方法研究,哈尔滨工程大学博士学位论文,2008
[40]刘畅,改进的HMM训练方法在语音识别中的应用,吉林大学硕士学位论文,2007.
[41]丁琼,网络化智能控制系统研究,贵州大学硕士学位论文,2008
[42]赵洪华,输油管线破坏预防监测及其定位技术的研究,济南大学硕士学位论文,2006
[43]张耀兰,语音识别技术在导航设备中的应用,北京交通大学硕士学位论文,2009
[44]张志霞,语音识别中个人特征参数提取研究,中北大学硕士学位论文,2009
[45]王一平用遗传算法改进HMM的语音识别算法研究太原理工大学硕士学位论文,2007
[46]刘涛,管道检测系统中的信号处理技术研究,天津大学硕士学位论文,2007
[47]徐毅琼,人脸识别技术研究,中国人民解放军信息工程大学硕士学位论文,2005
[48]张志刚,基于神经网络/HMM的语音识别算法的研究,武汉理工大学硕士学位论文,2006
[49]李建宁,汉语孤立词语音识别的研究与实现,西北大学硕士学位论文,2007
[50]项静恬等,多种时序模型的建模比较,数理统计与管理,1991年2月
[51] George,G.S.Forecasting chaotic time series with genetic algorithms. Physical Review E,1997,55(3)
[52]袁俊. HMM连续语音识别中Viterbi算法的优化及应用.电子技术[J]2001年第28卷第期48-51。