hmm和神经网络用于语音识别的算法研究-计算机应用技术专业毕业论文-research on algorithms of hmm and neural network for speech recognition - graduation thesis of computer application technology major.docx

太原理工大学硕士研究生学位论文论文又研究了几种可用于分类的前馈神经网络：反向传播网络 (BP网)，径向基函数(RBF)网络及小波神经网络，讨论了 各自的原理、学习过程以及用于语音识别的建模方法。RBF网 络隐节点的中心选取对网络性能有很大影响，常用的K．均值聚 类是一种无监督的学习方法，论文提出利用训练样本中的分类 信息对输入数据进行聚类，计算它们的形心作为隐节点函数中 一11,。实验结果表明这种有监督地选取函数中一11,识别率明显高于 K．均值聚类的方法。最后论文引入小波变换理论，用小波基函 数代替RBF网络中的高斯基函数构成一个小波RBF网络，实验 结果表明，这种神经网络同样可以获得很好的识别效果。关键词：语音识别，特征提取，隐马尔可夫模型，RBF神经网络，小 波神经网络太原理工大学硕士研究生学位论文STUDY OF SPEECH RECOGNITION ALGORITHM BASED oN HMM AND NEURAL NETWORKABSTRACTSpeech recognition is the research hotspot in the自ekl of speech sig芏1al processing．It haS been a difflcult problem for a long time，especially for the recognition of personImdependent and in noisy environment．This paper discussed several common speechrecogIlitiOn methods including classical Hidden Markov Model andartificial neural n酡ⅣOrk which is very popular curremly．It also introduced a new antimoise feature parameterZem．crossings wimPe a：k_眦plitlldes feature(ZCPA fjmlre)，which can be used toconstmct a robust speech rccognitiOn system．This p印er presentl甜several fhmiliar feature extracting methods such as Linear Prediction Cepstnlm Coef五cient(LPCC) and Mel Frequency Cepstmm Coemcient(MFCC)．They have got excellent recognition results under clean envir∞ment，but也eir perfbmance will deteriorate severely in noisy condition．So most pan is deVo￡ed to introduce ZCPA fbature and analyze its anti．noise principle．Then this paper discussed}丑、／IM theo巧which is used inspeech recognition and its jmp】ement撕on process．There areIII太原理工大学硕士研究生学位论文underflow problems in software implementation procedure for the classical Baum-Welch training algorithm，and a lot of literatures did not presented an explicit method．With respect to this problem，thispaper inducted the scaling algorithm and derived the reestimate formulae of Baum-Welch algorithm again．The experiments showed that it can converge rapidly and the recognition results are good， which proved the correctness of the new reestimate formulae，while the old formulae can not converge in experiments．Then the paper