基于语音反演机器学习方法的声道模型分析

《基于语音反演机器学习方法的声道模型分析》由会员上传分享，免费在线阅读，更多相关内容在应用文档-天天文库。

1、AbstractTheperformanceoftheautomaticspeechrecognition(ASR)systemsisaffectedbecauseofcoarticulation.Existingrelatedstudieshaveclaimedthatarticulatoryinformationcanbeusedtoimprovetheperformanceofautomaticspeechrecognitionsystems.However,sucharticulatoryinformationisnotsoeasytobeobt

2、ainedintypicalspeaker-listenersituations.Thatiswhytheacoustic-to-articulatoryspeechinversionisproposed.Acoustic-to-articulatoryspeechinversion(speechinversion)isamethodofestimatingarticulatorytrajectoriesorvocaltractconfigurationsfromthespeechsignal.Ifarticulatoryinformationcanbe

3、estimatedaccurately,itwillbeusefulforspeechsynthesis,languageacquisition,speechvisualizationandsoon.Firstly,tractvariables(insteadoftraditionalpellettrajectories)areusedasarticulatoryinformationtomodelspeechdynamicsandtheestimationperformanceandnon-uniquenessoftractvariablesandpe

4、llettrajectoriesarecomparedinthispaper.Thespeechsignalsareparameterizedasmel-frequencycepstralcoefficients(MFCC),perceptuallinearpredictioncepstralcoefficients(PLPCC)andlinearpredictioncepstralcoefficients(LPCC),andmixturedensitynetworks(MDN)isusedtoestimatetractvariablesandpelle

5、ttrajectories.Theresultsindicatethattractvariablescanprovideabetterestimationperformancethanpellettrajectories.Furthermore,amodel-basedstatisticalparadigmisusedtocalculatetheNormalizedNon-Uniqueness(NNU)andtheresultsshowthatuniquenessintheTV-basedinversemodeliscomparativelylowert

6、hanthepellet-basedmodelforthesamesixconsonants.Secondly,fourdifferentmachinelearningmethodsareusedforspeechinversion,whicharefeedforwardartificialneuralnetwork(FF-ANN),autoregressiveartificialneuralnetwork(AR-ANN),distalsupervisedlearning(DSL)andtrajectorymixturedensitynetwork(TM

7、DN),tocomparetractvariablesandpellettrajectories.Theresultsindicatethattractvariableshavebetterperformancethanpellettrajectoriesandaremorefitforarticulatoryfeature-basedASRsystems.Inaddition,theestimationperformanceofthesemachinelearningmethodsfortractvariableswhenspeechsignalisp

8、arameterizedasMFCCandacousticparameters(