《机器人语音识别中英文对照外文翻译文献.docx》由会员分享，可在线阅读，更多相关《机器人语音识别中英文对照外文翻译文献.docx（19页珍藏版）》请在taowenge.com淘文阁网|工程机械CAD图纸|机械工程制图|CAD装配图下载|SolidWorks_CaTia_CAD_UG_PROE_设计图分享下载上搜索。
1、 外文资料翻译中英文资料外文翻译译文:改进型智能机器人的语音识别方法2、语音识别概述最近，由于其重大的理论意义和实用价值，语音识别已经受到越来越多的关注。到现在为止，多数的语音识别是基于传统的线性系统理论，例如隐马尔可夫模型和动态时间规整技术。随着语音识别的深度研究，研究者发现，语音信号是一个复杂的非线性过程，如果语音识别研究想要获得突破，那么就必须引进非线性系统理论方法。最近，随着非线性系统理论的发展，如人工神经网络，混沌与分形，可能应用这些理论到语音识别中。因此，本文的研究是在神经网络和混沌与分形理论的基础上介绍了语音识别的过程。语音识别可以划分为独立发声式和非独立发声式两种。非独立发声式
2、是指发音模式是由单个人来进行训练，其对训练人命令的识别速度很快，但它对与其他人的指令识别速度很慢，或者不能识别。独立发声式是指其发音模式是由不同年龄， 不同性别，不同地域的人来进行训练，它能识别一个群体的指令。一般地，由于用户不需要操作训练，独立发声式系统得到了更广泛的应用。 所以，在独立发声式系统中，从语音信号中提取语音特征是语音识别系统的一个基本问题。语音识别包括训练和识别，我们可以把它看做一种模式化的识别任务。通常地， 语音信号可以看作为一段通过隐马尔可夫模型来表征的时间序列。通过这些特征提取，语音信号被转化为特征向量并把它作为一种意见，在训练程序中，这些意见将反馈到 HMM 的模型参数
3、估计中。这些参数包括意见和他们响应状态所对应的概率密度函数，状态间的转移概率，等等。经过参数估计以后，这个已训练模式就可以应用到识别任务当中。输入信号将会被确认为造成词，其精确度是可以评估的。整个过程如图一所示。3、理论与方法图 1 语音识别系统的模块图从语音信号中进行独立扬声器的特征提取是语音识别系统中的一个基本问题。解决这个问题的最流行方法是应用线性预测倒谱系数和 Mel 频率倒谱系数。这两种方法都是基于一种假设的线形程序， 该假设认为说话者所拥有的语音特性是由于声道共振造成的。这些信号特征构成了语音信号最基本的光谱结构。然而，在语音信号中，这些非线形信息不容易被当前的特征提取逻辑方法所提
4、取，所以我们使用分型维数来测量非线形语音扰动。本文利用传统的LPCC 和非线性多尺度分形维数特征提取研究并实现语音识别系统。3.1 线性预测倒谱系数线性预测系数是一个我们在做语音的线形预分析时得到的参数，它是关于毗邻语音样本间特征联系的参数。线形预分析正式基于以下几个概念建立起来的，即一个语音样本可以通过一些以前的样本的线形组合来快速地估计，根据真实语音样本在确切的分析框架（短时间内的）和预测样本之间的差别的最小平方原则，最后会确认出唯一的一组预测系数。LPC 可以用来估计语音信号的倒谱。在语音信号的短时倒谱分析中，这是一种特殊的处理方法。信道模型的系统函数可以通过如下的线形预分析来得到：其中
5、 p 代表线形预测命令， ，（k=1，2， ，p）代表预测参数，脉冲响应用h(n)来表示，假设 h（n）的倒谱是。那么（1）式可以扩展为（2）式：将（ 1）带入（ 2），两边同时 ，（2）变成（ 3）。就获得了方程（ 4）：那么 可以通过 来获得。（5）中计算的倒谱系数叫做 LPCC，n 代表 LPCC 命令。在我们采集 LPCC 参数以前，我们应该对语音信号进行预加重，帧处理，加工和终端窗口检测等，所以，中文命令字“前进”的端点检测如图 2 所示，接下来，断点检测后的中文命令字“前进”语音波形和 LPCC 的参数波形如图 3 所示。图 2 中文命令字“前进”的端点检测图 3 断点检测后的中文
6、命令字“前进”语音波形和 LPCC 的参数波形3.2 语音分形维数计算分形维数是一个与分形的规模与数量相关的定值，也是对自我的结构相似性的测量。分形分维测量是6-7。从测量的角度来看，分形维数从整数扩展到了分数，打破了一般集拓扑学方面被整数分形维数的限制，分数大多是在欧几里得几何尺寸的延伸。有许多关于分形维数的定义，例如相似维度，豪斯多夫维度，信息维度，相关维度，容积维度，计盒维度等等，其中，豪斯多夫维度是最古老同时也是最重要的， 它的定义如【3】所示：其中，表示需要多少个单位 来覆盖子集 F.端点检测后，中文命令词“向前”的语音波形和分形维数波形如图 4 所示。图 4 端点检测后，中文命令词
7、“向前”的语音波形和分形维数波形3.3 改进的特征提取方法考虑到 LPCC 语音信号和分形维数在表达上各自的优点，我们把它们二者混合到信号的特取中，即分形维数表表征语音时间波形图的自相似性，周期性，随机性， 同时，LPCC 特性在高语音质量和高识别速度上做得很好。由于人工神经网络的非线性，自适应性，强大的自学能力这些明显的优点，它的优良分类和输入输出响应能力都使它非常适合解决语音识别问题。由于人工神经网络的输入码的数量是固定的，因此，现在是进行正规化的特征参数输入到前神经网络9，在我们的实验中，LPCC 和每个样本的分形维数需要分别地通过时间规整化的网络，LPCC 是一个 4 帧数据（LPCC
8、1,LPCC2,LPCC3,LPCC4，每个参数都是 14 维的），分形维数被模范化为 12 维数据，（FD1,FD2, FD12，每一个参数都是一维），以便于每个样本的特征向量有 4*14+12*1=68-D 维，该命令就是前56 个维数是 LPCC，剩下的 12 个维数是分形维数。因而，这样的一个特征向量可以表征语音信号的线形和非线性特征。自动语音识别的结构和特征自动语音识别是一项尖端技术，它允许一台计算机，甚至是一台手持掌上电脑（迈尔斯，2000）来识别那些需要朗读或者任何录音设备发音的词汇。自动语音识别技术的最终目的是让那些不论词汇量，背景噪音，说话者变音的人直白地说出的单词能够达到
9、100%的准确率（CSLU，2002）。然而，大多数的自动语音识别工程师都承认这样一个现状，即对于一个大的语音词汇单位，当前的准确度水平仍然低于90%。举一个例子，Dragons Naturally Speaking 或者 IBM 公司，阐述了取决于口音，背景噪音，说话方式的基线识别的准确性仅仅为 60%至 80%(Ehsani & Knodt, 1998)。更多的能超越以上两个的昂贵的系统有 Subarashii (Bernstein, et al., 1999), EduSpeak (Franco, etal., 2001), Phonepass (Hinks, 2001), ISLE P
10、roject (Menzel, et al., 2001) and RAD (CSLU, 2003)。语音识别的准确性将有望改善。在自动语音识别产品中的几种语音识别方式中，隐马尔可夫模型（ HMM）被认为是最主要的算法，并且被证明在处理大词汇语音时是最高效的(Ehsani & Knodt, 1998)。详细说明隐马尔可夫模型如何工作超出了本文的范围，但可以在任何关于语言处理的文章中找到。其中最好的是 Jurafsky & Martin (2000) and Hosom, Cole, and Fanty (2003)。简而言之，隐马尔可夫模型计算输入接收信号和包含于一个拥有数以百计的本土音素录音
11、的数据库的匹配可能性 (Hinks, 2003, p. 5)。也就是说，一台基于隐马尔可夫模型的语音识别器可以计算输入一个发音的音素可以和一个基于概率论相应的模型达到的达到的接近度。高性能就意味着优良的发音，低性能就意味着劣质的发音(Larocca, et al., 1991)。虽然语音识别已被普遍用于商业听写和获取特殊需要等目的，近年来，语言学习的市场占有率急剧增加(Aist, 1999; Eskenazi, 1999; Hinks, 2003)。早期的基于自动语音识别的软件程序采用基于模板的识别系统，其使用动态规划执行模式匹配或其他时间规范化技术(Dalby & Kewley-Port,1
12、999). 这些程序包括 Talk to Me (Auralog, 1995), the Tell Me More Series (Auralog, 2000), Triple-Play Plus (Mackey & Choi, 1998), New Dynamic English (DynEd, 1997), English Discoveries (Edusoft, 1998), and See it, Hear It, SAY IT! (CPI, 1997) 。这些程序的大多数都不会提供任何反馈给超出简单说明的发音准确率，这个基于最接近模式匹配说明是由用户提出书面对话选择的。学习者不会被
13、告之他们发音的准确率。特别是内里，（2002 年）评论例如 Talk to Me 和 Tell Me More 等作品中的波形图，因为他们期待浮华的买家，而不会提供有意义的反馈给用户。Talk to Me 2002 年的版本已经包含了更多 Hinks (2003)的特性，比如，信任对于学习者来说是非常有用的: 一个视觉信号可以让学习者把他们的语调同模型扬声器发出的语调进行对比。 学习者发音的准确度通常以数字 7 来度量（越高越好） 那些发音失真的词语会被识别出来并被明显地标注。原文:Improved speech recognition method for intelligent robot
14、2、Overview of speech recognitionSpeech recognition has received more and more attention recently due the important theoretical meaning and practical value 5 . Up to now, m speechrecognitioinsbased on conventionalinearsystem theory,such asHidden Markov Model (HMM) and Dynamic Time Warping(DTW) . With
15、 the deep studyof speech recognitioni,tis found thatspeech signalis acomplex nonlinear process. If the study of speech recognition wants to b through, nonlinear-system theory method must be introducedto it.Recently,with the developmentof nonlinea-systemtheories suchasartificianleural networks(ANN) ,
16、 chaos and fractal, it is possible to apply these theorie speech recognition. Therefore, the study of this paper is based on ANN a chaos and fractal theories are introduced to process speech recognition.Speech recognition is divided into two ways that are speaker dependen and speakerindependent.Spea
17、ker dependentrefersto the pronunciationmodel trainedby a singleperson,the identificatiroanteof the training person?sorderisshigh,while othersordeirssin low identificatiroanteorcant be recognizSepde.aker independent refers to the pronunciation modeltrained by persons of different age, sex and region,
18、 it can identify a g personsordeGresn.erallys,peaker independenstystem ismorewidelyused,sincethe user is not requiredto conduct the trainingS.o extractionof speakerindependentfeaturesfrom the speech signalis the fundamental problem of speaker recognition system.Speech recognition can be viewed as a
19、pattern recognition task, which includes training and recognition.Generally, speech signal can be viewed a time sequenceand characterizebdy thepowerfulhidden Markov model(HMM). Through the feature extraction, the speech signal is transferred featurevectorsand act asobservationsI.n the trainingproced
20、ure,these observationswilfleed to estimatethe model parametersof HMM.These parametersincludeprobabilitdyensityfunctionfor the observationsand theircorrespondingstatest,ransitiopnrobabilitbyetween the statese,tc.After the parameter estimation,the trainedmodels can be used for recognition task. The in
21、put observations will be recognized as the resul words and the accuracy can be evaluated. Thewhole process is illustrated Fig. 1.Fig. 1 Block diagram of speech recognition system3Theory andmethodExtraction of speaker independent features from the speech signal is fundamental problem of speaker recog
22、nitionsystem. Thestandard methodology for solvingthisproblem uses Linear PredictiveCepstral Coefficient(sLPCC) and Mel-Frequency CepstralCo-efficien(tMFCC).Both thesemethods are linearproceduresbased on the assumptionthatspeaker features have properties caused by the vocal tract resonances. T featur
23、es form the basic spectral structure of the speech signal. However non-linearinformationin speech signalsis not easilyextractedby the presentfeatureextractiomnethodologies.So we use fractadlimensiontomeasure non2linear speech turbulence.This paper investigateasnd implementsspeakeridentificatisoynste
24、m usingboth traditionaLlPCC and non-linearmultiscaledfractadlimension feature extraction.3. 1 L inear Predictive Cepstral CoefficientsLinear predictioncoefficien(tLPC) is a parameter setwhich isobtained when we do linear prediction analysis of speech. It is about so correlationcharacteristicbsetween
25、 adjacent speech samples. Linear predictionanalysisis based on the followingbasicconcepts.That is,aspeech sample can be estimated approximately by the linear combination o some past speech samples. According to the minimal square sum principle of differencebetween real speech sample in certainanalys
26、isframeshort-timeand predictivseample,the only group ofpredictiocnoefficients can be determined.LPC coefficient can be used to estimate speech signal cepstrum. This a special processing method in analysis of speech signal short-time ceps System function of channelmodel is obtained by linear predicti
27、on analysi follow.Where p representlsinearpredictioonrder,ak,(k=1,2,pr)epresent spredictiocnoefficienItm,pulse responseisrepresentedby h(n).Supposecepstrum of h(n) is represented by,then (1) can be expanded as (2).The cepstrum coefficient calculated in the way of (5) is called LPCC represents LPCC o
28、rder.When we extract LPCC parameter before, we should carry on speech signal pre-emphasis,framing processing,windowingprocessing and endpoints detectioentc. , so the endpoint detection of Chinese command word“Forward”sihsown in Fig.2,next,the speech waveform ofChinese commandword“Forward”anLdPCCpara
29、meterwaveform afterEndpoint detection is shown in Fig. 3.3. 2 Speech Fractal Dimension ComputationFractal dimension is a quantitative value from the scale relati on the meaning of fractala,nd alsoa measuringon self-similariotfyits structureT.hefractalmeasuring is fractaldimension6-7.Fromthe viewpoin
30、tof measuring,fractaldimension is extended from integerto fractionb,reakingthe limitofthe generalto pology setdimension being integerFractaldimension,fractiomnostly, is dimension extensioninEuclidean geometry.There are many definitions on fractal dimension, eg.,similar dimensi Hausdoff dimension, in
31、foration dimension, correlation dimension, capabilimension, box-counting dimension etc. , where,Hausdoff dimension is oldest and also most important, for any sets, it is defined as3.Where, M(F) denotes how many unit needed to cover subset F.In thispaper, the Box-Counting dimension (DB) of ,F, is obt
32、ained by partitionintghe plane with squaresgridsof side , and the numberofsquares that intersect the pla)n)ea(nNd(is defined as8.The speech waveform of Chinese commandword“Forward”and fractal dimension waveform after Endpoint detection is shown in Fig. 4.3. 3 Improved feature extractions methodConsi
33、deringthe respectivaedvantageson expressingspeech signalof LPCC and fractal dimension,we mix both to be the feature signal, that fractal dimension denotes the self2similarity, periodicity and randomnes speech time wave shape,meanwhile LPCC featureis good for speechquality and high on identification
34、rate.Due to ANNs nonlinearitsye,l-fadaptabilitryo,bustand self-learningsuch obvious advantages, its good classification and input2output reflec ability are suitable to resolve speech recognition problem.Due to the number of ANNinputnodes being fixed,thereforetime regularization is carried out to the
35、 feature parameter before inputted t neural network9. In our experiments, LPCC and fractal dimension of eacsample are need to get through the network of time regularization separa LPCCis 4-frame data(LPCC1,LPCC2,LPCC3,LPCC4,eachframeparameter is 14-D), fractaldimension is regularizedto be12-frame da
36、ta(FD1,FD2,FD12,each frame parameteris1-D),so thatthe feature vectorof each sample has 4*14+1*12=68-D, the order is,the first56dimensions are LPCC, the rest 12 dimensions are fractal dimensions. Thus such mixed featureparameter can showspeech linearand nonlinear characteristics as well.Architectures
37、 and Features of ASRASRis a cuttingedge technologythatallowsa computer or even a hand-heldPDA(Myers, 2000) to identifywords thatare read aloud or spoken intoany sound-recordingdevice.The ultimatepurpose of ASR technologyisto allow 100%accuracywith allwords thatare intelligibly spoken byany personreg
38、ardlesosf vocabularysize, backgrounndoise,orspeaker variables (CSLU, 2002). However, most ASR engineers admit that thecurrentaccuracylevelfora largevocabularyunitof speech (e.g.t,he sentence)remains lessthan 90%. DragonsNaturallySpeaking or IBMs ViaVoice, for example, show a baseline recognition acc
39、uracy of only 60% 80%, depending upon accent,background noise,type of utterancee,tc. (Ehsani&Knodt, 1998). More expensivesystems thatare reportedto outperformthesetwo are Subarashii(Bernsteine,t al.,1999),EduSpeak(Franco, et al., 2001), Phonepass (Hinks, 2001), ISLE Project (Menzel, e 2001) and RAD
40、(CSLU, 2003). ASR accuracy is expected to improve.Among several types of speech recognizers used in ASR products, both implemented and proposed, the Hidden Markov Model (HMM) is one of the most dominant algorithmsand has proven to be an effectivmeethod of dealingwith largeunitsof speech (Ehsani &Kno
41、dt, 1998). Detailed descriptionosf how the HHMmodel works go beyond the scope of thispaper and can be found in any textconcernedwith languageprocessing;among the best are Jurafsky & Martin (2000) and Hosom, Cole, and Fanty(2003). Put simply, HMM computes the probable match between the input it recei
42、vesand phonemes containedin a databaseof hundreds of nativespeaker recordings (Hinks, 2003, p. 5). That is, a speech recognizer bas HMMcomputes howclosethe phonemes of a spoken inputare to a correspondingmodel, based onprobabilitytheory.Highlikelihood represents good pronunciation; low likelihood re
43、presents poor pronunciat (Larocca, et al., 1991).While ASRhas been commonly used for such purposesas business dictatioannd specialneeds accessibilitiyt,smarket presenceforlanguage learninghas increaseddramaticalliyn recentyears(Aist,1999; Eskenazi, 1999; Hinks, 2003). Early ASR-basedsoftware program
44、s adopted template-baserdecognitiosnystemswhich performpatternmatching using dynamic programming or othertime normalizationtechniques(Dalby & Kewley-Port,1999). These programisncludeTalk to Me (Auralog,1995),the Tell Me More Series (Auralog, 2000), Triple-Play Plus (Mackey & Choi 1998), New Dynamic
45、English (DynEd, 1997), English Discoveries (Edusoft, 1998), and See it, Hear It, SAY IT! (CPI, 1997). Most of these programs not provide any feedback on pronunciationaccuracy beyond simply indicatinwghich writtendialoguechoicethe userhas made, based on the closestpatternmatch. Learners are not told
46、the accuracy of their pronunciation. In particular, Neri, et al. (2002) criticizes the graphic forms presented in products such as Talk to Me and TellbeMceauMsoeretheylook flashyto buyers,but do not givemeaningfulfeedbackto users.The 2000 version of Talk to Me has incorporated more of the features t
47、ha Hinks (2003), for example, believes are useful to learners: A visual signal allows learners to compare their intonation to that ofmodel speaker. The learners pronunciation accuracy is scored on a scale of seven (thhigher the better).Words whose pronunciation fails to be recognized are highlighted