Noise-estimation-by-minima-controlled-recursive-av

12IEEESIGNALPROCESSINGLETTERS,VOL.9,NO.1,JANUARY2002NoiseEstimationbyMinimaControlledRecursiveAveragingforRobustSpeechEnhancementIsraelCohen,Member,IEEE,andBaruchBerdugoAbstract—Inthisletter,weintroduceaminimacontrolledrecursiveaveraging(MCRA)approachfornoiseestimation.Thenoiseestimateisgivenbyaveragingpastspectralpowervaluesandusingasmoothingparameterthatisadjustedbythesignalpresenceprobabilityinsubbands.Presenceofspeechinsubbandsisdeterminedbytheratiobetweenthelocalenergyofthenoisyspeechanditsminimumwithinaspecifiedtimewindow.Thenoiseestimateiscomputationallyefficient,robustwithrespecttotheinputsignal-to-noiseratio(SNR)andtypeofunderlyingadditivenoise,andcharacterizedbytheabilitytoquicklyfollowabruptchangesinthenoisespectrum.IndexTerms—Acousticnoise,signaldetection,spectralanalysis,speechenhancement.I.INTRODUCTIONACRUCIALcomponentofapracticalspeechenhancementsystemistheestimationofthenoisepowerspectrum.Acommonapproachistoaveragethenoisysignalovernonspeechsections.Aspeechpausedetectioniseitherimplementedonaframe-by-framebasis[1]orestimatedindependentlyforindi-vidualsubbandsusingaposteriorisignal-to-noiseratio(SNR)[2],[3].However,thedetectionreliabilityseverelydeterioratesforweekspeechcomponentsandlow-inputSNR.Additionally,theamountofpresumablenonspeechsectionsinthesignalmaynotbesufficient,whichrestrictsthetrackingcapabilityofthenoiseestimatorincaseofvaryingnoisespectrum.Alternatively,thenoisecanbeestimatedfromhistogramsinthepowerspectraldomain[3]–[5].Unfortunately,suchmethodsarecomputation-allyexpensive.Martin[6]hasproposedanalgorithmfornoiseestimationbasedonminimumstatistics.Thenoiseestimateisobtainedastheminimavaluesofasmoothedpowerestimateofthenoisysignal,multipliedbyafactorthatcompensatesthebias.How-ever,thisnoiseestimateissensitivetooutliers[5]anditsvari-anceisabouttwiceaslargeasthevarianceofaconventionalnoiseestimator[6].Moreover,thismethodmayoccasionallyattenuatelowenergyphonemes,particularlyiftheminimumsearchwindowistooshort[7].Acomputationallymoreeffi-cientminimumtrackingschemeispresentedin[8].ItsmaindrawbackistheveryslowupdaterateofthenoiseestimateinManuscriptreceivedFebruary22,2001;October26,2001.Theassociateed-itorcoordinatingthereviewofthismanuscriptandapprovingitforpublicationwasDr.A.S.Spanias.I.CoheniswiththeDepartmentofElectricalEngineering,Technion—IsraelInstituteofTechnology,Haifa32000,Israel(e-mail:[e-mail]).B.BerdugoiswithLamarSignalProcessing,Ltd.,YokneamIlit20692,Israel(e-mail:[e-mail]).PublisherItemIdentifierS1070-9908(02)02410-0.caseofasuddenriseinnoiseenergylevelanditstendencytocancelthesignal[9].Inthisletter,weintroduceaminimacontrolledrecursiveav-eraging(MCRA)approachfornoiseestimation.Thenoisees-timateisgivenbyaveragingpastspectralpowervalues,usingasmoothingparameterthatisadjustedbythesignalpresenceprobabilityinsubbands.Weshowthatpresenceofspeechinagivenframeofasubbandcanbedeterminedbytheratiobe-tweenthelocalenergyofthenoisyspeechanditsminimumwithinaspecifiedtimewindow.Theratioiscomparedtoacer-tainthresholdvalue,whereasmallerratioindicatesabsenceofspeech.Subsequently,atemporalsmoothingiscarriedouttoreducefluctuationsbetweenspeechandnonspeechsegments,therebyexploitingthestrongcorrelationofspeechpresenceinneighboringframes.Theresultantnoiseestimateiscomputa-tionallyefficient,robustwithrespecttotheinputSNRandtypeofunderlyingadditivenoiseandcharacterizedbytheabilitytoquicklyfollowabruptchangesinthenoisespectrum.Theletterisorganizedasfollows.InSectionII,wepresentthenoisespectrumestimationapproach.InSectionIII,wein-troduceaminimacontrolledestimatorforthespeechpresenceprobability.InSectionIV,weevaluatetheproposedmethodanddiscussexperimentalresults,whichvalidateitsusefulness.II.NOISESPECTRUMESTIMATIONLetanddenotespeechanduncorrelatedadditivenoisesignals,respectively,whereisadiscrete-timeindex.Theobservedsignal,givenby,isdividedintooverlappingframesbytheapplicationofawindowfunctionandanalyzedusingtheshort-timeFouriertransform(STFT).Specifically,(1)whereisthefrequencybinindex,isthetimeframeindex,isananalysiswindowofsize,andistheframeupdatestepintime.Giventwohypotheses,and,whichindicate,respectively,speechabsenceandpresenceinthethframeofthethsubband,wehave(2)whereandrepresenttheSTFTofthecleanandnoisesignals,respectively.Letdenotethevarianceofthenoiseinthethsubband.Thenacommon1070–9908/02$17.00©2002IEEECOHENANDBERDUGO:NOISEESTIMATIONROBUSTSPEECHENHANCEMENT13techniquetoobtainitsestimateistoapplyatemporalrecursivesmoothingtothenoisymeasurementduringperiodsofspeechabsence.Inparticular,(3)whereisasmoothingparameterandanddesignatehypotheticalspeechabsenceandpresence,respec-tively.Here,wemakeadistinctionbetweenthehypothesesin(2),usedforestimatingthecleanspeechandthehypothesesin(3),whichcontroltheadaptationofthenoisespectrum.Clearly,decidingspeechisabsent()whenspeechispresent()ismoredestructivewhenestimatingthesignalthanwhenesti-matingthenoise.Hence,differentdecisionrulesareemployedandgenerallywetendtodecidewithahigherconfidencethan,i.e.,[7].Letdenotetheconditionalsigna