用人工神经网络的不同方法消除语音信号中的噪声(IJITCS-V7-N7-2)

I.J.InformationTechnologyandComputerScience,2015,07,8-18PublishedOnlineJune2015inMECS()DOI:10.5815/ijitcs.2015.07.02Copyright©2015MECSI.J.InformationTechnologyandComputerScience,2015,07,8-18RemovingNoisefromSpeechSignalsUsingDifferentApproachesofArtificialNeuralNetworksOmaimaN.A.AL-AllafFacultyofSciences&IT,Dept.ofBasicSciences,Al-ZaytoonahUniversityofJordan,P.O.Box130,Amman(11733),JordanE-mail:[e-mail]—Inthisresearch,fourANNmodels:FunctionFitting(FitNet),NonlinearAutoRegressive(NARX),Recurrent(RNNs),andCascaded-ForwardNetwereconstructedandtrainedseparatelytobecomeafiltertoremovenoisefromanyspeechsignal.Eachmodelconsistsofinput,hiddenandoutputlayers.Twoneuronsintheinputlayerthatrepresentspeechsignalanditsassociatednoise.Theoutputlayerincludesoneneuronthatrepresenttheenhancedsignalafterremovingnoise.Thefourmodelsweretrainedseparatelyonstereo(noisyandclean)audiosignalstoproducethecleansignal.Experimentswereconductedforeachmodelseparatelywithdifferent:architecture;optimizationtrainingalgorithms;andlearningparameterstoidentifymodelwithbestresultsofremovingnoisefromspeechsignal.Fromexperiments,bestresultswereobtainedfromFitNetandNARAXmodelsrespectively.TrainLMisthebesttrainingalgorithminthiscase.Finally,theresultsshowedthatthesuggestedarchitectureofthefourmodelshavefilteringabilitytoremovenoiseformbothtrainedandnottrainedspeechsignalssamples.IndexTerms—SignalEnhancement,ArtificialNeuralNetworks,FunctionFitting(FitNet),NonlinearAutoRegressive(NARX),Recurrent(RNNs),andCascaded-ForwardNetI.INTRODUCTIONArtificialNeuralNetworks(ANNs)includesmanyneuronsthatoperateinparallelandconnectedviaweights[1].ANNshavebeenusedlargelyinthefieldsofimageandsignalprocessing[2].ManyANNmodelsthatdifferinarchitectureandtrainingalgorithmsweresuggestedinthepast.Multi-layerperceptronandbackpropagationneuralnetwork(BPNN)arelargelyusedANNmodels.Thesemodelsrequirelargenumberofiterationsfortrainingnetworkonanyspecifiedproblem.Manyliteratureapproacheshavebeencarriedouttoimprovethespeedupthenetworklearning[3][4][28..32].ThenonlinearANNnatureandabilitytolearnfromtheirenvironmentsinsupervisedandunsupervisedwaysmakethemhighlysuitedforsolvingdifficultsignalprocessingproblems.Butalltheseavailabletechniqueshavecertainlimitations.Signalenhancementisprocessofperformingnonlinearfilteringofasignalfornoisereduction.Iscriticaltounderstandthenatureofproblemformulationwhendealingwithsignalprocessingapplications,sothatthemostANNapproachescanbeapplied.Recently,ANNsarefoundtobeaveryefficienttoolforsignalenhancement.ItisimportanttoassesstheimpactofANNontheperformance,robustness,andcost-effectivenessofsignalprocessingsystems.AnotherimportantissueishowtoevaluateANNapproaches,learningalgorithms,andstructuresforsolvingsignalprocessingproblems[5].ANNsmayprovideanewapproachforsignalfiltering.RecentexperimentalsuggeststhatANNsmaybeabletoreducenoiseformanyapplications[6].TherearefewliteratureresearchesadopteddifferentANNsmodelsforremovingnoisefromsignal.Atthesametime,thefilteringabilityofANNarchitecturesofmostoftheseliteraturestudieswerelimitedtotrainingsetsignals.Theseliteratureapproacheswerenotabletoremovenoisefromuntrainedsignalsorfilteredthenoisewithbadqualityandgeneratethesignalwithsomenoise.Otherliteraturestudiesrequiredlongtimefortrainingprocess.Therefore,weneedANNapproachesandmodelsforsignalenhancementthatremovenoisefromanyspeechsignal.Also,weneedANNmodelsthatgeneratehighvaluesofPeakSignaltoNoiseRatio(PSNR)andlessvaluesofMeanSquareError(MSE).Weneedtoreducethetimerequiredforthetrainingprocess.Inthisresearch,fourANNmodelswereconstructedandimplemented(inMathLab2013a)forsignalenhancementsystem:FunctionFitting(FitNet),NonlinearAutoregressive(NARX),Recurrent(RNNs),andCascaded-ForwardNettoimprovethespeedofconvergenceandenhancethesignalsbyobtaininghighPSNR,MSEandbestvaluesofR2(coefficientofdeterminationofalinearregression).Threeoptimizationalgorithmswereusedintrainingprocess(LM,GD,andGDM).ManyexperimentswereconductedtodeterminetheANNmodelwithitsarchitecturethatleadtolesslearningtimeandbestperformance.Comparisonsbetweenthefourmodelswereconductedalso.Theresearchisorganizedasfollows:sectionIIincludesrelatedliteratureandsectionIIIdescribessignalenhancement.SectionIVincludesdetailsaboutANNusedmodels.SectionVincludesresearchmethodology.SectionVIincludesexperimentsandfinallysectionVIIconcludesthiswork.RemovingNoisefromSpeechSignalsUsingDifferentApproachesofArtificialNeuralNetworks9Copyright©2015MECSI.J.InformationTechnologyandComputerScience,2015,07,8-18II.RELATEDLITERATUREManyliteraturestudiesfornoisereductionfromspeechsignalwerebasedondifferentapproaches.OtherliteraturestudieswerebasedonANNforsignalenhancementeachwithdifferentmodelandarchitecture.Thegoalistodefineamodelthatgivebestresultsforsignalenhancement.Atthebeginning,Kevin,1988[6]exploredBPNNapproachtofilteringnoisefromsignals.Experimentswereconductedusingsinglesinewaveinputs,multiplesinewaveinputs,andhumanspeechinputs.Thenetworks'outputswerethencomparedtotheoriginalsign