說(shuō)話人聲音轉(zhuǎn)換方法的研究.pdf

1、中國(guó)科學(xué)技術(shù)大學(xué)碩士學(xué)位論文說(shuō)話人聲音轉(zhuǎn)換方法的研究姓名：王子祥申請(qǐng)學(xué)位級(jí)別：碩士專業(yè)：信號(hào)與信息處理指導(dǎo)教師：戴禮榮20050501中周科學(xué)技術(shù)大學(xué)填f‘論文AbstractAbstractSpeechsignalsconveyawiderangeofinformationAmongthem，themeaningofthemessagecaniedbyspeechsignalisofprimeimportanceMeanwhile，t

2、hesecondaryinformationcardedbyspeechsignalsuchasspeakeridentityalsoplaysanimportantroleinhumancommunicationVoiceconversiontechniquesattempttomodifyspeechsignalSOthatitisperceivedasifspokenbyanotherspeaker(targetspeaker)W

3、ithvariousspeechsystemsbecomingmoreandmoreubiquitous，voiceconversiontechniquescanofferanumberofusefulapplicationsForexample，anovicetLsercalladaptatexttospeechsystemtosynthesizespeechwithanotherspeakeridentityquicklyandin

4、expensivelybasedOilvoiceconversiontechniqueTherearetwoclassicalgroupsofmethodsforvoiceconversion；thosebyGMMandthosebycodebookmappingButthesetwomethodstendtogenerateoverlysmoothedeffectonconvertedspeechspectrumSotheconver

5、tedspeechqualityisdecreasedgreatlyToovercomethisproblem，altercarefullyunderstandingthemethodofcodebookmapping，alotofexperimentshavebeendoneonconvertedspeechspectrum，Andanovelvoiceconversionmethodbasedoncodebookmappingwit

6、hphonemetiedweightingissummarizedinthisdissertationbyanalyzingresultsoftheexperimentsExperimentsshowthatwiththismetllIodmuchimprovementhasbeenmadeinoverlysmoothedeffectonconvertedspeechspectrum，andthequalityofconvertedsp

7、eechhasbeenimprovedalotMeanwhile，aprosodicconversionmethodbydecisiontreeisappliedtocopewiththecomplexprosodyofChineseThesubjectivetestsshowthattheproposedmethodscalleffectivelyCOnVertspeakersindividualitywhilemaintaining

8、highspeechqualitywithonlyasmallamountoftrainingdataInrecentyearstheconceptofTrainable”ShasbeenproposedandwidelyusedInsuchaTTSsystem，speechsynthesisunitsaremodeledbymultispaceprobabilitydistributionHMM，whichcanmodefspeech

9、spectrumandpitchsimultaneouslyinaunifiedframeworkAnewvoiceconversiontechniquebasedonTrainableTTSframeworkisalsopresentedinthisdissertationComparedwiththeclassicalmethod，thistechniqueaimstoadapttheHMMofsourcespeakertothat

10、oftargetspeakerusingadaptationtechniqueSpeechparametersusedinthesynthesisstagearegenerateddirectlyfromadaptedHMMbyusingaparametergenerationalgorithmPrimaryExperimentsshowthat，using10adaptationsentences，speechsamplesgener