資源描述:
《基于gmm獨(dú)立建模語音轉(zhuǎn)換系統(tǒng)的研究》由會(huì)員上傳分享,免費(fèi)在線閱讀,更多相關(guān)內(nèi)容在學(xué)術(shù)論文-天天文庫。
1、基于GMM的獨(dú)立建模語音轉(zhuǎn)換系統(tǒng)研究中文摘要語音轉(zhuǎn)換就是對(duì)一個(gè)說話人(源說話人)的語音信號(hào)進(jìn)行轉(zhuǎn)換,使之聽起來像另一個(gè)說話人(目標(biāo)說話人)語音的技術(shù)。這項(xiàng)技術(shù)幾乎囊括了語音信號(hào)處理領(lǐng)域的各個(gè)方面,它的研究對(duì)語音分析,語音編碼,語音合成,語音增強(qiáng),語音識(shí)別等方面有重要的促進(jìn)作用。本文提出了一種基于GMM獨(dú)立建模的轉(zhuǎn)換方法,主要內(nèi)容包括:(1)提出了一種獨(dú)立建模的方法,對(duì)源和目標(biāo)說話人語音特征分別建立GMM模型,解決了傳統(tǒng)基于GMM的聯(lián)合建模轉(zhuǎn)換系統(tǒng)中,需要并行語料,不適應(yīng)多人轉(zhuǎn)換等缺點(diǎn)。-(2)通過輸入語音特征矢量,動(dòng)態(tài)確定轉(zhuǎn)換規(guī)則,進(jìn)行語音轉(zhuǎn)換,解
2、決了傳統(tǒng)轉(zhuǎn)換系統(tǒng)中,轉(zhuǎn)換函數(shù)非動(dòng)態(tài)這一缺陷,使得轉(zhuǎn)換規(guī)則更為靈活,從而增強(qiáng)轉(zhuǎn)換精確性。(3)基于線性預(yù)鋇JJ(LP)分析合成平臺(tái),對(duì)源與目標(biāo)說話人LPCC特征獨(dú)立建立GMM模型,對(duì)測(cè)試語音進(jìn)行轉(zhuǎn)換,分析得到轉(zhuǎn)換結(jié)果,并做出相應(yīng)評(píng)測(cè)與分析。(4)基于STRAIGHT分析合成平臺(tái),對(duì)源與目標(biāo)說話人STRAIGHT分析得到的譜包絡(luò)進(jìn)行建模分析,并對(duì)測(cè)試語音進(jìn)行轉(zhuǎn)換,得到結(jié)果,并做出評(píng)測(cè)分析。+(5)基于語音結(jié)構(gòu)化(AUS)理論,改進(jìn)轉(zhuǎn)換系統(tǒng),成功實(shí)現(xiàn)語音轉(zhuǎn)換,解決了在轉(zhuǎn)換系統(tǒng)中,源與目標(biāo)相同音素聚類的GMM分量如何對(duì)齊這一難題,通過系統(tǒng)實(shí)驗(yàn),取得了階段性
3、成果。(6)以基于GMM的源與目標(biāo)聯(lián)合建模的模型為藍(lán)本,比較獨(dú)立建模的轉(zhuǎn)換系統(tǒng),相對(duì)于同樣的測(cè)試語音,其結(jié)果的差異性。關(guān)鍵字:語音轉(zhuǎn)換,獨(dú)立建模,GMM作者:徐小峰指導(dǎo)老師:俞一彪TheResearchoficeConve"S);toBasedlCe0nversionSystemase0nGMMwithIndependentSpeakermodelingAbstractVoiceconversionischangingonespeaker’S(sourcespeaker)acousticfeaturestoanother(targetspeaker
4、),thenitwasheardasifutteredbytargetspeaker.Itcontainsatmostallthebranchesofspeechsignalprocessingarea,sotheresearchisusefulforspeechcodec,speechsynthesis,speechenhancement,speechrecognition,etc.inthisdissertation,anindependentspeakermodelingvoiceconversionalgorithmbasedonGMMis
5、proposed,themainworkasfollows:(1)proposedanewmethodbasedonindependentspeakermodeling,sourceandtargetacousticfeaturesismodeledindependentbyGMM,thisalgorithmnotonlyavoidthedisadvantagesoftraditionalmethodthatneedsparallelvoicedatabase,butalsoreducecomplexdegreeofthesystem,especi
6、allyinmulti—peopleconversionsystem..(2)thesystemadoptdynamicrulesintransformationprocess,whichadjustrulesreal-timeaccordingtoinputspeechdata,SOthespectrumconversionismoreaccuratethangeneralmethod..(3)chooselinearpredictionsynthesisplatform,thenuseGMMtrainingsourceandtargetlpcc
7、features,attheend,transformthetestspeech,analysistheresultsandgivetheSCOre.(4)chooseSTRAIGHTsynthesisplatform,thenuseSTRAIGHTtogetspectrumenvelope,thenuseGMMtrainingSOurCeandtargetfeatures,transformthetestspeech,analysistheresultandgivethescore.(5)useAUStheorytoimprovesystempe
8、rformance,solutetheproblemthathowtoaligntwomodelcomponents,It