基于多特征有效組合的說(shuō)話人識(shí)別

《基于多特征有效組合的說(shuō)話人識(shí)別》由會(huì)員上傳分享，免費(fèi)在線閱讀，更多相關(guān)內(nèi)容在教育資源-天天文庫(kù)。

1、基于多特征有效組合的說(shuō)話人識(shí)別墊查董!圣王壁查組僉的說(shuō)話人識(shí)別基于多特征有效組合的說(shuō)話人識(shí)別謝迎春,于湘珍,劉建平.,張衛(wèi)華(1.武警工程學(xué)院研究生隊(duì)陜西西安710086;2.武警工程學(xué)院電子技術(shù)基礎(chǔ)實(shí)驗(yàn)室陜西西安710086;3.武警工程學(xué)院通信工程系無(wú)線通信工程教研室陜西西安710086;4.武警工程學(xué)院通信工程系信息工程教研室陜西西安710086)摘要:通過(guò)分析"-3今說(shuō)話人識(shí)別系統(tǒng)中常用的一些特征參數(shù),以提高說(shuō)話人識(shí)別的識(shí)別率為目的,在Matlab6.5軟件環(huán)境下提出了將Mel頻率倒譜(MFCC),線性預(yù)測(cè)倒譜(LPCC)及他們的一階差分和基音周期等多種特征有效結(jié)合進(jìn)行說(shuō)話

2、人識(shí)別的方法采用短時(shí)自相關(guān)法提取基音周期,在識(shí)別過(guò)程中采用改進(jìn)的動(dòng)態(tài)規(guī)整算法,將模板的匹配過(guò)程與檢驗(yàn)量的計(jì)算分離開(kāi),每幀給出一個(gè)說(shuō)話人辨認(rèn)結(jié)果,最后綜合各幀的辨認(rèn)結(jié)果,得出最佳匹配結(jié)果.經(jīng)過(guò)多次實(shí)驗(yàn)證明,采用以上方法使用多特征有效結(jié)合比單個(gè)使用各種特征效果要好,能在一定程度上提高系統(tǒng)區(qū)分說(shuō)話人的能力.關(guān)鍵詞:說(shuō)話人識(shí)別;動(dòng)態(tài)規(guī)整;MFCC;LPCC;基音周期中圖分類號(hào):TN912文獻(xiàn)標(biāo)識(shí)碼:B文章編號(hào):1004—373X(2005)09—068—03SpeakerIdentificationBasedonEfficientlyCombiningManifoldFeaturesXIE

3、Yingchun,YUXiangzhen,LIUJianping.,ZHANGWeihua(1.GraduateStudentTeamtCollegeofArmedPoliceForcetXian,710086.China;2.ElectronTechniqueBasicLab,CollegeofArmedPoliceForcetxiant710086tChina;3.TeachingChamberofWirelessCommunicationEngineeringinCommunicationEngineeringDepartmenttCollegeofArmedPoliceFo

4、rce,Xiant710086,China4.TeachingChamberofComputerInformationProjectinCommunicationEngineeringDepartment,CollegeofArmedPoliceForce.Xian,710086.China)Abstract:Throughanalyzingsomefeaturesthatbeusedusuallyinspeakeridentificationsystemnowadays,inordertOimprovetherateofidentification,thispaperputsfo

5、rwardamethodthatcombiningefficientlymorefeaturessuchasMFCCandLPCCandtheironerankscoefficientsandkeynoteperiodandSOontodospeakerverificationunderMatlab6.5.Wepickupkeynoteperiodsbyself—correlationmethodanduseanewDynamicTimeWarping(DTW)methodtodoidentification.ThisnewDTWmethodisawaythatdividingte

6、mplatematchingandcalculationoftestmeasureandcalculatingidentificationresultsofallframesaftereveryframegettingoutaidentificationresult.Atlast,wecanmakeoutthebestmatchingresult.Throughseriesofexperiments,itprovesthatthemethodofusingmanifoldfeaturesisbetterthanthemethodofusingsinglefeatureandthea

7、bilityofspeakeridentificationcanbeimprovedbyusingthisway.Keywords:speakerverification;DTW;MFCC;LPCC;keynoteperiod1引言說(shuō)話人識(shí)別是語(yǔ)音識(shí)別的一個(gè)分支,在公安偵察,聲控系統(tǒng),醫(yī)療診斷,電子金融業(yè)務(wù)等方面有著廣泛的應(yīng)用前景.他和語(yǔ)音識(shí)別的區(qū)別在于,他并不注意語(yǔ)音信號(hào)中的語(yǔ)義內(nèi)容,而是希望從語(yǔ)音信號(hào)中提取出個(gè)人的信息特征.從這點(diǎn)上說(shuō),說(shuō)話人識(shí)別是謀求挖掘出包含在