SIST EN 300 969 V7.0.1:2003

2003-01.Slovenski inštitut za standardizacijo. Razmnoževanje celote ali delov tega standarda ni dovoljeno.Digital cellular telecommunications system (Phase 2+) (GSM); Half rate speech; Half rate speech transcoding (GSM 06.20 version 7.0.1 Release 1998)33.070.50Globalni sistem za mobilno telekomunikacijo (GSM)Global System for Mobile Communication (GSM)ICS:Ta slovenski standard je istoveten z:EN 300 969 Version 7.0.1SIST EN 300 969 V7.0.1:2003en01-december-2003SIST EN 300 969 V7.0.1:2003SLOVENSKI
STANDARD
ETSIEN300969V7.0.1(2000-01)EuropeanStandard(Telecommunicationsseries)Digitalcellulartelecommunicationssystem(Phase2+);Halfratespeech;Halfratespeechtranscoding(GSM06.20version7.0.1Release1998)GLOBALSYSTEMFORMOBILECOMMUNICATIONSRSIST EN 300 969 V7.0.1:2003

ETSIETSIEN300969V7.0.1(2000-01)2(GSM06.20version7.0.1Release1998)ReferenceREN/SMG-110620Q7KeywordsDigitalcellulartelecommunicationssystem,GlobalSystemforMobilecommunications(GSM),CODEC,GSM,speechETSIPostaladdressF-06921SophiaAntipolisCedex-FRANCEOfficeaddress650RoutedesLucioles-SophiaAntipolisValbonne-FRANCETel.:+33492944200Fax:+33493654716SiretN°34862356200017-NAF742CAssociationàbutnonlucratifenregistréeàlaSous-PréfecturedeGrasse(06)N°7803/88Internetsecretariat@etsi.frIndividualcopiesofthisETSIdeliverablecanbedownloadedfromhttp://www.etsi.orgIfyoufinderrorsinthepresentdocument,sendyourcommentto:editor@etsi.frImportantnoticeThisETSIdeliverablemaybemadeavailableinmorethanoneelectronicversionorinprint.Inanycaseofexistingorperceiveddifferenceincontentsbetweensuchversions,thereferenceversionisthePortableDocumentFormat(PDF).Incaseofdispute,thereferenceshallbetheprintingonETSIprintersofthePDFversionkeptonaspecificnetworkdrivewithinETSISecretariat.CopyrightNotificationNopartmaybereproducedexceptasauthorizedbywrittenpermission.Thecopyrightandtheforegoingrestrictionextendtoreproductioninallmedia.©EuropeanTelecommunicationsStandardsInstitute2000.Allrightsreserved.SIST EN 300 969 V7.0.1:2003

ETSIETSIEN300969V7.0.1(2000-01)3(GSM06.20version7.0.1Release1998)ContentsIntellectualPropertyRights.5Foreword.51Scope.72References.73Definitions,symbolsandabbreviations.73.1Definitions.73.2Symbols.93.3Abbreviations.104FunctionaldescriptionoftheGSMhalfratespeechcodec.114.1GSMhalfratespeechencoder.114.1.1High-passfilter.134.1.2Segmentation.144.1.3FixedPointLatticeTechnique(FLAT).144.1.4Spectralquantization.154.1.4.1AutocorrelationFixedPointLatticeTechnique(AFLAT).154.1.5Frameenergycalculationandquantization.174.1.6Softinterpolationofthespectralparameters.174.1.7Spectralnoiseweightingfiltercoefficients.184.1.8LongTermPredictorlagdetermination.194.1.8.1Openlooplongtermsearchinitialization.204.1.8.2Openlooplagsearch.214.1.8.3Framelagtrajectorysearch(Mode≠0).264.1.8.4Voicingmodeselection.284.1.8.5Closedlooplagsearch.284.1.9Harmonicnoiseweighting.294.1.10Codesearchalgorithm.314.1.10.1Decorrelationoffilteredbasisvectors.324.1.10.2Fastsearchtechnique.334.1.11Multimodegainvectorquantization.344.1.11.1CodingGSandP0.344.2GSMhalfratespeechdecoder.374.2.1Excitationgeneration.374.2.2Adaptivepitchprefilter.384.2.3SynthesisFilter.384.2.4Adaptivespectralpostfilter.384.2.5Updatingdecoderstates.405Homingsequences.405.1Functionaldescription.405.2Definitions.405.3Encoderhoming.405.4Decoderhoming.415.5Encoderhomestate.415.6Decoderhomestate.41AnnexA(normative):Codecparameterdescription.42A.1Codecparameterdescription.42A.1.1MODE.42A.1.2R0.42A.1.3LPC1-LPC3.43A.1.4LAG_1-LAG_4.43A.1.5CODEx_1-CODEx_4.43A.1.6GSP0_1-GSP0_4.43SIST EN 300 969 V7.0.1:2003

ETSIETSIEN300969V7.0.1(2000-01)4(GSM06.20version7.0.1Release1998)A.2Basiccoderparameters.43AnnexB(normative):OrderofoccurrenceofthecodecparametersoverAbis.44AnnexC(informative):Bibliography.45AnnexD(informative):ChangeRequestHistory.46History.47SIST EN 300 969 V7.0.1:2003

ETSIETSIEN300969V7.0.1(2000-01)5(GSM06.20version7.0.1Release1998)IntellectualPropertyRightsIPRsessentialorpotentiallyessentialtothepresentdocumentmayhavebeendeclaredtoETSI.TheinformationpertainingtotheseessentialIPRs,ifany,ispubliclyavailableforETSImembersandnon-members,andcanbefoundinSR000314:"IntellectualPropertyRights(IPRs);Essential,orpotentiallyEssential,IPRsnotifiedtoETSIinrespectofETSIstandards",whichisavailablefromtheETSISecretariat.LatestupdatesareavailableontheETSIWebserver(http://www.etsi.org/ipr).PursuanttotheETSIIPRPolicy,noinvestigation,includingIPRsearches,hasbeencarriedoutbyETSI.NoguaranteecanbegivenastotheexistenceofotherIPRsnotreferencedinSR000314(ortheupdatesontheETSIWebserver)whichare,ormaybe,ormaybecome,essentialtothepresentdocument.ForewordThisEuropeanStandard(Telecommunicationsseries)hasbeenproducedbytheSpecialMobileGroup(SMG).ThepresentdocumentspecifiesthespeechcodectobeusedfortheGSMhalfratechannelforthedigitalcellulartelecommunicationssystem.Thepresentdocumentispartofaseriescoveringthehalfratespeechtrafficchannelsasdescribedbelow:GSM06.02"Digitalcellulartelecommunicationssystem(Phase2+);Halfratespeech;Halfratespeechprocessingfunctions".GSM06.06"Digitalcellulartelecommunicationssystem(Phase2+);Halfratespeech;ANSI-CcodefortheGSMhalfratespeechcodec".GSM06.07"Digitalcellulartelecommunicationssystem(Phase2+);Halfratespeech;TestsequencesfortheGSMhalfratespeechcodec".GSM06.20"Digitalcellulartelecommunicationssystem(Phase2+);Halfratespeech;Halfratespeechtranscoding".GSM06.21"Digitalcellulartelecommunicationssystem(Phase2+);Halfratespeech;Substitutionandmutingoflostframesforhalfratespeechtrafficchannels".GSM06.22"Digitalcellulartelecommunicationssystem(Phase2+);Halfratespeech;Comfortnoiseaspectsforhalfratespeechtrafficchannels".GSM06.41"Digitalcellulartelecommunicationssystem(Phase2+);Halfratespeech;DiscontinuousTransmission(DTX)forhalfratespeechtrafficchannels".GSM06.42"Digitalcellulartelecommunicationssystem(Phase2+);Halfratespeech;VoiceActivityDetector(VAD)forhalfratespeechtrafficchannels".ThecontentsofthepresentdocumentissubjecttocontinuingworkwithinSMGandmaychangefollowingformalSMGapproval.ShouldSMGmodifythecontentsofthepresentdocumentitwillbere-releasedwithanidentifyingchangeofreleasedateandanincreaseinversionnumberasfollows:Version7.x.ywhere:7indicatesRelease1998ofGSMPhase2+.xtheseconddigitisincrementedforallchangesofsubstance,i.e.technicalenhancements,corrections,updates,etc.ythethirddigitisincrementedwheneditorialonlychangeshavebeenincorporatedinthespecification.SIST EN 300 969 V7.0.1:2003

ETSIETSIEN300969V7.0.1(2000-01)6(GSM06.20version7.0.1Release1998)NationaltranspositiondatesDateofadoptionofthisEN:31December1999DateoflatestannouncementofthisEN(doa):31March2000DateoflatestpublicationofnewNationalStandardorendorsementofthisEN(dop/e):30September2000DateofwithdrawalofanyconflictingNationalStandard(dow):30September2000SIST EN 300 969 V7.0.1:2003

ETSIETSIEN300969V7.0.1(2000-01)7(GSM06.20version7.0.1Release1998)1ScopeThepresentdocumentspecifiesthespeechcodectobeusedfortheGSMhalfratechannel.Italsospecifiesthetestmethodstobeusedtoverifythatthecodecimplementationcomplieswiththepresentdocument.TherequirementsaremandatoryforthecodectobeusedeitherinGSMMobileStations(MS)sorBaseStationSystems(BSS)sthatutilizethehalfrateGSMspeechtrafficchannel.2ReferencesThefollowingdocumentscontainprovisionswhich,throughreferenceinthistext,constituteprovisionsofthepresentdocument.•Referencesareeitherspecific(identifiedbydateofpublication,editionnumber,versionnumber,etc.)ornon-specific.•Foraspecificreference,subsequentrevisionsdonotapply.•Foranon-specificreference,thelatestversionapplies.•Anon-specificreferencetoanETSshallalsobetakentorefertolaterversionspublishedasanENwiththesamenumber.•ForthisRelease1998document,referencestoGSMdocumentsareforRelease1998versions(version7.x.y).[1]GSM06.02:"Digitalcellulartelecommunicationssystem(Phase2+);Halfratespeech;Halfratespeechprocessingfunctions".[2]GSM06.06:"Digitalcellulartelecommunicationssystem(Phase2+);Halfratespeech;ANSI-CcodefortheGSMhalfratespeechcodec".[3]GSM06.07:"Digitalcellulartelecommunicationssystem(Phase2+);Halfratespeech;TestsequencesfortheGSMhalfratespeechcodec".3Definitions,symbolsandabbreviations3.1DefinitionsForthepurposesofthepresentdocument,thefollowingtermsanddefinitionsapply.adaptivecodebook:Theadaptivecodebookisderivedfromthelongtermfilterstate.Thelagvaluecanbeviewedasanindexintotheadaptivecodebook.adaptivepitchprefilter:IntheGSMhalfratespeechdecoder,thisfilterisappliedtotheexcitationsignaltoenhancetheperiodicityofthereconstructedspeech.Notethatthisisdonepriortotheapplicationoftheshorttermfilter.adaptivespectralpostfilter:IntheGSMhalfratespeechdecoder,thisfilterisappliedtotheoutputoftheshorttermfiltertoenhancetheperceptualqualityofthereconstructedspeech.allowablelags:ThesetoflagvalueswhichmaybecodedbytheGSMhalfratespeechencoderandtransmittedtotheGSMhalfratespeechdecoder.Thissetcontainsbothintegerandfractionalvalues(seetable3).analysiswindow:Foreachframe,theshorttermfiltercoefficientsarecomputedusingthehighpassfilteredspeechsampleswithintheanalysiswindow.Theanalysiswindowis170samplesinlength,andiscenteredaboutthelast100samplesintheframe.basisvectors:AsetofM,M1,orM2vectorsoflengthNsusedtogeneratetheVSELPcodebookvectors.Thesevectorsarenotnecessarilyorthogonal.SIST EN 300 969 V7.0.1:2003

ETSIETSIEN300969V7.0.1(2000-01)8(GSM06.20version7.0.1Release1998)closedlooplagsearch:Aprocessofdeterminingthenearoptimallagvaluefromtheweightedinputspeechandthelongtermfilterstate.closedlooplagtrajectory:Foragivenframe,thesequenceofnearoptimallagvalueswhoseelementscorrespondtoeachofthefoursubframesasdeterminedbytheclosedlooplagsearch.codebook:Asetofvectorsusedinavectorquantizer.Codeword(ORCode):AnM,M1,orM2bitsymbolindicatingthevectortobeselectedfromaVSELPcodebook.Delta(LAG)code:Afourbitcodeindicatingthechangeinlagvalueforasubframerelativetotheprevioussubframe'scodedlag.Forframesinwhichthelongtermpredictorisenabled(MODE1,2,or3),thelagforsubframe1isindependentlycodedusingeightbits,anddeltacodesareusedforsubframes2,3,and4.directformcoefficients:Oneoftheformatsforstoringtheshorttermfilterparameters.Allfilterswhichareusedtomodifyspeechsamplesusedirectformcoefficients.fractionallags:Asetoflagvalueshavingsub-sampleresolution.NotethatnoteveryfractionallagvalueconsideredintheGSMhalfratespeechencoderisanallowablelagvalue.frame:Atimeintervalequalto20ms,or160samplesatan8kHzsamplingrate.harmonicnoiseweightingfilter:Thisfilterexploitsthenoisemaskingpropertiesofthespectralpeakswhichoccuratharmonicsofthepitchfrequencybyweightingtheresidualerrorlessinregionsnearthepitchharmonicsandmoreinregionsawayfromthem.Notethatthisfilterisonlyusedwhenthelongtermfilterisenabled(MODE=1,2or3).highpassfilter:Thisfilterisusedtode-emphasizethelowfrequencycomponentsoftheinputspeechsignal.integerlags:Asetoflagvalueshavingwholesampleresolution.interpolatingfilter:AnFIRfilterusedtoestimatesub-sampleresolutionsamples,givenaninputsampledwithintegersampleresolution.lag:Thelongtermfilterdelay.Thisistypicallythepitchperiod,oramultipleorsub-multipleofit.longtermfilter:Thisfilterisusedtogeneratetheperiodiccomponentintheexcitationforthecurrentsubframe.ThisfilterisonlyenabledforMODE=1,2or3.LPCcoefficients:LinearPredictiveCoding(LPC)coefficientsisagenericdescriptivetermfordescribingtheshorttermfiltercoefficients.openlooplagsearch:Aprocessofestimatingthenearoptimallagdirectlyfromtheweightedspeechinput.Thisisdonetonarrowtherangeoflagvaluesoverwhichtheclosedlooplagsearchshallbeperformed.openlooplagtrajectory:Foragivenframe,thesequenceofnearoptimallagvalueswhoseelementscorrespondtothefoursubframesasdeterminedbytheopenlooplagsearch.reflectioncoefficients:Analternativerepresentationoftheinformationcontainedintheshorttermfilterparameters.residual:Theoutputsignalresultingfromaninversefilteringoperation.shorttermfilter:Thisfilterintroduces,intotheexcitationsignal,shorttermcorrelationwhichmodelstheimpulseresponseofthevocaltract.softinterpolation:Aprocesswhereinadecisionismadeforeachframetouseeitherinterpolatedoruninterpolatedshorttermfilterparametersforthefoursubframesinthatframe.softinterpolationbit:Aonebitcodeindicatingwhetherornotinterpolationoftheshorttermparametersistobeusedinthecurrentframe.spectralnoiseweightingfilter:Thisfilterexploitsthenoisemaskingpropertiesoftheformants(vocaltractresonances)byweightingtheresidualerrorlessinregionsneartheformantfrequenciesandmoreinregionsawayfromthem.subframe:Atimeintervalequalto5ms,or40samplesatan8kHzsamplingrate.SIST EN 300 969 V7.0.1:2003

ETSIETSIEN300969V7.0.1(2000-01)9(GSM06.20version7.0.1Release1998)vectorquantization:Amethodofgroupingseveralparametersintoavectorandquantizingthemsimultaneously.GSP0vectorquantizer:Theprocessofvectorquantization,itsintermediateparameters(GSandP0)forthecodingoftheexcitationgainsβandγ.VSELPcodebook:Vector-SumExcitedLinearPredictive(VSELP)codebook,usedintheGSMhalfratespeechcoder,whereineachcodebookvectorisconstructedasalinearcombinationofthefixedbasisvectors.zeroinputresponse:Theoutputofafilterduetoallpastinputs,i.e.duetothepresentstateofthefilter,giventhataninputofzerosisapplied.zerostateresponse:Theoutputofafilterduetothepresentinput,giventhatnopastinputshavebeenapplied,i.e.giventhestateinformationinthefilterisallzeroes.3.2SymbolsForthepurposesofthepresentdocument,thefollowingsymbolsapply:A(z)Shorttermspectralfilter.αiTheLPCcoefficients.bL(n)Theoutputofthelongtermfilterstate(adaptivecodebook)forlagL.βThelongtermfiltercoefficient.C(z)Secondweightingfilter.e(n)Weightederrorsignalfj(i)Thecoefficientsofthejthphaseofthe10thorderinterpolatingfilterusedtoevaluatecandidatefractionallagvalues;irangesfrom0toPf-1.gj(i)Thecoefficientsofthejthphaseofthe6thorderinterpolatingfilterusedtointerpolateC'sandG'saswellasfractionallagsintheharmonicnoiseweighting;irangesfrom0toPg-1.γThegainappliedtothevector(s)selectedfromtheVSELPcodebook(s).HAM2bitcodeindicatingthevectortobeselectedfromthesecondVSELPcodebook(whenoperatinginmode0).IAMorM1bitcodeindicatingthevectortobeselectedfromoneofthetwofirstVSELPcodebooks.LThelongtermfilterlagvalue.Lmax142(samples),themaximumpossiblevalueforthelongtermfilterlag.Lmin21(samples),theminimumpossiblevalueforthelongtermfilterlag.M9,thenumberofbasisvectors,andthenumberofbitsinacodeword,fortheVSELPcodebookusedinmodes1,2,and3.M17,thenumberofbasisvectors,andthenumberofbitsinacodeword,forthefirstVSELPcodebookusedinmode0.M27,thenumberofbasisvectors,andthenumberofbitsinacodeword,forthesecondVSELPcodebookusedinmode0.MODEAtwobitcodeindicatingthemodeforthecurrentframe(seeannexA).NA170,thelengthoftheanalysiswindow.Thisisthenumberofhighpassfilteredspeechsamplesusedtocomputetheshorttermfilterparametersforeachframe.NF160,thenumberofsamplesperframe(atasamplingrateof8kHz).Np10,theshorttermfilterorder.Ns40,thenumberofsamplespersubframe(atasamplingrateof8kHz).P16,thenumberofbitsintheprequantizerforther1-r3vectorquantizer.P25,thenumberofbitsintheprequantizerforther4-r6vectorquantizer.P34,thenumberofbitsintheprequantizerforther7-r10vectorquantizer.PfTheorderofonephaseofaninterpolatingfilterusedtoevaluatecandidatefractionallagvalues.Pfequals10forj≠0andequalto1forj=0.PgTheorderofonephaseofaninterpolatingfilter,fj(n),usedtointerpolateC'sandG'saswellasfractionallagsintheharmonicnoiseweighting,Pgequals6.pitchThetimedurationbetweentheglottalpulseswhichresultwhenthevocalchordsvibrateduringspeechproduction.Q111,thenumberofbitsinther1-r3reflectioncoefficientvectorquantizer.Q29,thenumberofbitsinther4-r6reflectioncoefficientvectorquantizer.Q38,thenumberofbitsinther7-r10reflectioncoefficientvectorquantizer.SIST EN 300 969 V7.0.1:2003

ETSIETSIEN300969V7.0.1(2000-01)10(GSM06.20version7.0.1Release1998)R0Afivebitcodeusedtoindicatetheenergylevelinthecurrentframe.r(n)Thelongtermfilterstate(thehistoryoftheexcitationsignal);n<0rL(n)ThelongtermfilterstatewiththeadaptivecodebookoutputforlagLappended.s'(n)Synthesizedspeech.W(z)Spectralweightingfilter.λhnwTheharmonicnoiseweightingfiltercoefficient.ξTheadaptivepitchprefiltercoefficient.xCeilingfunction:thelargestintegerywherey
ETSIETSIEN300969V7.0.1(2000-01)11(GSM06.20version7.0.1Release1998)4FunctionaldescriptionoftheGSMhalfratespeechcodecTheGSMhalfratecodecusestheVSELP(Vector-SumExcitedLinearPrediction)algorithm.TheVSELPalgorithmisananalysis-by-synthesiscodingtechniqueandbelongstotheclassofspeechcodingalgorithmsknownasCELP(CodeExcitedLinearPrediction).TheGSMhalfratecodec'sencodingprocessisperformedona20msspeechframeatatime.Aspeechframeofthesampledspeechwaveformisreadandbasedonthecurrentwaveformandthepasthistoryofthewaveform,thecodecencoderderives18parametersthatdescribeit.Theparametersextractedaregroupedintothefollowingthreegeneralclasses:-energyparameters(R0andGSP0);-spectralparameters(LPCandINT_LPC);-excitationparameters(LAGandCODE).Theseparametersarequantizedinto112bitsfortransmissionasdescribedinannexAandtheirorderofoccurrenceoverAbisisgiveninannexB.TheGSMhalfratecodecisananalysis-by-synthesiscodec,thereforethespeechdecoderisprimarilyasubsetofthespeechencoder.Thequantizedparametersaredecodedandasyntheticexcitationisgeneratedusingtheenergyandexcitationparameters.Thesyntheticexcitationisthenfilteredtoprovidethespectralinformationresultinginthegenerationofthesynthesizedspeech(seefigure1).GSMhalfratespeechcodecSpeechencoderSpeechdecoderspeechtransmittedspeechparametersreceivedspeechparameterssynthesisedspeechFigure1:BlockdiagramoftheGSMhalfratespeechcodecTheANSI-CcodethatdescribestheGSMhalfratespeechcodecisgiveninGSM06.06[2]andthetestsequencesinGSM06.07[3](seeclause5forthecodechomingtestsequences).4.1GSMhalfratespeechencoderTheGSMhalfratespeechencoderusesananalysisbysynthesisapproachtodeterminethecodetousetorepresenttheexcitationforeachsubframe.ThecodebooksearchprocedureconsistsoftryingeachcodevectorasapossibleexcitationfortheCodeExcitedLinearPredictive(CELP)synthesizer.Thesynthesizedspeechs'(n)iscomparedagainsttheinputspeechandadifferencesignalisgenerated.Thisdifferencesignalisthenfilteredbyaspectralweightingfilter,W(z),(andpossiblyasecondweightingfilter,C(z))togenerateaweightederrorsignal,e(n).Thepowerine(n)iscomputed.Thecodevectorwhichgeneratestheminimumweightederrorpowerischosenasthecodevectorforthatsubframe.Thespectralweightingfilterservestoweighttheerrorspectrumbasedonperceptualconsiderations.Thisweightingfilterisafunctionofthespeechspectrumandcanbeexpressedintermsoftheαparametersoftheshortterm(spectral)filter.SIST EN 300 969 V7.0.1:2003

ETSIETSIEN300969V7.0.1(2000-01)12(GSM06.20version7.0.1Release1998)()WzzziiiNiiiNpp=−−−=−=1111αα~(1)Thecomputationoftheαicoefficientsisdescribedinsubclause4.1.7.ThesecondweightingfilterC(z),ifused,isaharmonicweightingfilterandisusedtocontroltheamountoferrorintheharmonicsofthespeechsignal.Iftheweightingfilter(s)aremovedtobothinputpathstothesubtracter,anequivalentconfigurationisobtainedasshowninfigure2.Σ( )2VSELPCodebookX_IγInputspeechs(n)y(n)y'(n)e(n)DetermineLPCcoefficientsβLtotalweightederrorLPC1LPC2LPC3R0INT_LPCMODELAG-1CODE_1LAG_2CODE_2LAG_3CODE_3LAG_4CODE_4GSP0-1GSP0_2GSP0_3GSP0_4W(z)C(z)FindMinimumoverLandallIB(z)H(z)C(z)βγfindoptimalgainsand.Figure2:BlockdiagramoftheGSMhalfratespeechencoder(MODE=1,2and3)HereH(z)isthecombinationofA(z),theshortterm(spectral)filter,andW(z),thespectralweightingfilter.ThesefiltersarecombinedsincethedenominatorofA(z)iscancelledbythenumeratorofW(z).SIST EN 300 969 V7.0.1:2003

ETSIETSIEN300969V7.0.1(2000-01)13(GSM06.20version7.0.1Release1998)()HzziiiNp=−−=111~α(2)Therearetwoapproachesthatcanbeusedforcalculatingthegain,γ.Thegaincanbedeterminedpriortocodebooksearchbasedonresidualenergy.Thisgainwouldthenbefixedforthecodebooksearch.Anotherapproachistooptimizethegainforeachcodevectorduringthecodebooksearch.Thecodevectorwhichyieldstheminimumweightederrorwouldbechosenanditscorrespondingoptimalgainwouldbeusedforγ.Thelatterapproachgenerallyyieldsbetterresultssincethegainisoptimizedforeachcodevector.Thisapproachalsoimpliesthatthegaintermneedstobeupdatedatthesubframerate.Theoptimalcodeandgainforthistechniquecanbecomputedasfollows:Theinputspeechisfirstfilteredbyahighpassfilterasdescribedinsubclause4.1.1.Theshorttermfilterparametersarecomputedfromthefilteredinputspeechonceperframe.Afastfixedpointcovariancelatticetechniqueisused.Subclauses4.1.3and4.1.4describesindetailhowtheshorttermparametersaredeterminedandquantized.Anoverallframeenergyisalsocomputedandcodedonceperframe.Onceperframe,oneofthefourvoicingmodesisselected.IfMODE≠0,thelongtermpredictorisusedandthelongtermpredictorlag,L,isupdatedatthesubframerate.LandaVSELPcodewordareselectedsequentially.Eachischosentominimizetheweightedmeansquareerror.Thelong-termfiltercoefficient,β,andthecodebookgain,γ,areoptimizedjointly.Subclause4.1.8describesthetechniqueforselectingfromamongthevoicingmodesand,ifoneofvoicedmodesischosen,determiningthelong-termfilterlag.Subclause4.1.10describesanefficienttechniqueforjointlyoptimizingβ,γandthecodewordselection.Subclause4.1.10alsoincludesthedescriptionofthefastVSELPcodebooksearchtechnique.Theβandγparametersaretransformedtoequivalentparametersusingtheframeenergyterm,andarevectorquantizedeverysubframe.Thecodingoftheframeenergyandtheβ andγparametersisdescribedinsubclause4.1.11.4.1.1High-passfilterThe13bitlinearPulseCodeModulated(PCM)inputspeech,x(n),isfilteredbyafourthorderpole-zerohighpassfilter.Thisfiltersuppressesthefrequencycomponentsoftheinputspeechwhicharebelow120Hz.Thefilterisimplementedasacascadeoftwosecond-orderInfiniteImpulseResponse(IIR)filters.Incorporatedintothefiltercoefficientsisagainof0,5.Thedifferenceequationforthefirstfilteris:()()()~~,,ynbxniaynjiijj=−+−==102112(3)where:b10=0,335052b11=-0,669983a11=0,926117b12=0,335052a12=-0,429413Thedifferenceequationforthesecondfilteris:()()()ynbyniaynjiijj=−+−==202212,,~(4)where:b20=0,335052b21=-0,669434a21=0,965332b22=0,335052a22=-0,469513SIST EN 300 969 V7.0.1:2003

ETSIETSIEN300969V7.0.1(2000-01)14(GSM06.20version7.0.1Release1998)4.1.2SegmentationAsamplebuffercontainingtheprevious195inputhighpassfilteredspeechsamples,y(n),isshiftedsothattheoldest160samplesareshiftedoutwhilethenext160inputsamplesareshiftedin.Theoldest160samplesinthebuffercorrespondtothenextframeofsamplestobeencoded.Theanalysisintervalcomprisesthemostrecent170samplesinthebuffer.Thesamplesinthebufferarelabelledass(n)where0≤n≤194ands(0)isthefirst(oldest)sample.4.1.3FixedPointLatticeTechnique(FLAT)Letrjrepresentthejthreflectioncoefficient.TheFLATalgorithmforthedeterminationofthereflectioncoefficientsisstatedasfollows:STEP1Computethecovariance(autocorrelation)matrixfromtheinputspeech:()()()φiksnisnknNNpA,=+−+−=24240≤i,k≤Np(5)STEP2Theφ(i,k)arrayismodifiedbywindowing()()()φφ',,ikikwik=−0≤i,k≤Np(6)STEP3()()Fikik0,',=φ0≤i,k≤Np-1(7)()()Bikik011,',=++φ0≤i,k≤Np-1(8)()()Cikik01,',=+φ0≤i,k≤Np-1(9)STEP4setj=1STEP5Computerj()()()()()()rCCNjNjFBFNjNjBNjNjjjjppjjjppjpp=−+−−++−−+−−−−−−−−2000000111111,,,,,,(10)STEP6Ifj=NPthendone.STEP7UpdateFj(i,k),Bj(i,k),Cj(i,k)0≤i,k≤NP-j-1()()()()()()FikFikrCikCkirBikjjjjjjj,,,,,=+++−−−−11121(11)()()()()()()BikBikrCikCkirFikjjjjjjj,,,,,=+++++++++++−−−−1112111111111,(12)()()()()()()CikCikrBikFikrCkijjjjjjj,,,,,=+++++++−−−−111211111(13)STEP8j=j+1STEP9gotostep5.Thewindowingcoefficients,w(|i-k|),arefoundinthetable1.SIST EN 300 969 V7.0.1:2003

ETSIETSIEN300969V7.0.1(2000-01)15(GSM06.20version7.0.1Release1998)Table1:Windowingcoefficientsw(0)0,998966w(5)0,974915w(1)0,996037w(6)0,969054w(2)0,991663w(7)0,963060w(3)0,986399w(8)0,956796w(4)0.980722w(9)0,950127Thisalgorithmcanbesimplifiedbynotingthattheφ',FandBmatricesaresymmetricsuchthatonlytheuppertriangularpartofthematricesneedtobecomputedorupdated.Also,step7isdonesothatFj(i,k),Bj(i-1,k-1),Cj(i,k-1),andCj(k,i-1)areupdatedtogetherandcommontermsarecomputedonceandtherecursionisdoneinplace.4.1.4SpectralquantizationAthreesegmentvectorquantizerofthereflectioncoefficientsisemployed.Areducedcomplexitysearchtechniqueisusedtoselectthevectorofreflectioncoefficientsforeachsegment.Thereflectioncoefficientvectorquantizercodebooksarestoredincompressedformtominimizetheirmemoryrequirements.Thethreesegmentsofthevectorquantizerspanreflectioncoefficientsr1r3,r4r6,andr7-r10respectively.Thebitallocationsforthevectorquantizersegmentsare:Q111bitsQ29bitsQ38bitsAreflectioncoefficientvectorprequantizerisusedateachsegment.Theprequantizersizeateachsegmentis:P16bitsP25bitsP34bitsAtagivensegment,theresidualerrorduetoeachvectorfromtheprequantizeriscomputedandstoredintemporarymemory.Thislistissearchedtoidentifythefourprequantizervectorswhichhavethelowestdistortion.Theindexofeachselectedprequantizervectorisusedtocalculateanoffsetintothevectorquantizertableatwhichthecontiguoussubsetofquantizervectorsassociatedwiththatprequantizervectorbegins.Thesizeofeachvectorquantizersubsetatthek-thsegmentisgivenby:SkQPkk=22(14)Thefoursubsetsofquantizervectors,associatedwiththeselectedprequantizervectors,aresearchedforthequantizervectorwhichyieldsthelowestresidualerror.Thusatthefirstsegment,64prequantizervectorsand128quantizervectorsareevaluated,32prequantizervectorsand64quantizervectorsareevaluatedatthesecondsegment,and16prequantizervectorsand64quantizervectorsareevaluatedatthethirdsegment.4.1.4.1AutocorrelationFixedPointLatticeTechnique(AFLAT)AnautocorrelationversionoftheFLATalgorithm,AFLAT,isusedtocomputetheresidualerrorenergyforareflectioncoefficientvectorbeingevaluated.ComputetheautocorrelationsequenceR(i),fromtheoptimalreflectioncoefficients,rj,overtherange0≤i≤Np.STEP1DefinetheinitialconditionsfortheAFLATrecursion:()()PiRiiNp001=≤≤−,(15)()()ViRiNiNpp0111=+−≤≤−,(16)SIST EN 300 969 V7.0.1:2003

ETSIETSIEN300969V7.0.1(2000-01)16(GSM06.20version7.0.1Release1998)STEP2Initializek,thevectorquantizersegmentindex:k=1(17)STEP3LetIl(k)betheindexofthefirstlatticestageinthek-thsegment,andIh(k)betheindexofthelastlatticestageinthek-thsegment.STEP4Initializej,theindexofthelatticestage,topointtothebeginningofthek-thsegment:()jIkl=(18)STEP5SettheinitialconditionsPj-1andVj-1to:()()()()PiPiiIkIkjjhl−−=≤≤−110,(19)()()()()()()ViViIkIkiIkIkjjihlhl−−=−+≤≤−1,(20)STEP6ComputethevaluesofVjandPjarraysusing:()()()()()[]()PirPirViViiIkjjjjjjjh=+++−≤≤−−−−−1012111,(21)()()()()ViVirVirPijNiNjjjijjijjpp=++−−+++−≤≤−−−−−11211121,(22)STEP7Incrementj:j=j+1STEP8Ifj
ETSIETSIEN300969V7.0.1(2000-01)17(GSM06.20version7.0.1Release1998)STEP14Ifj≤Ih(k)gotoSTEP12STEP15Incrementk,thevectorquantizersegmentindex:k=k+1STEP16Ifk≤3gotoSTEP4.Otherwise,theindicesofthereflectioncoefficientvectorsforthethreesegmentshavebeenchosen,andthesearchofthereflectioncoefficientvectorquantizeristerminated.Tominimizethestoragerequirementsforthereflectioncoefficientvectorquantizer,eightbitcodesfortheindividualreflectioncoefficientsarestoredinthevectorquantizertable,insteadoftheactualreflectioncoefficientvalues.Thecodesareusedtolookupthevaluesofthereflectioncoefficientsfromascalarquantizationtablewith256entries.4.1.5FrameenergycalculationandquantizationTheunquantizedvalueofR0,R(0),iscomputedduringthecomputationoftheshorttermpredictorparameters.()()()R01010320=+φφ0,0,(27)whereφ(i,k)isdefinedbyequation(5).R(0)isthenconvertedintodBrelativetofullscale(fullscale,Rmax,isdefinedasthesquareofthemaximumsampleamplitude).()RRRdB=10010logmax(28)RdBisthenquantizedto32levels.The32quantizedvaluesforRdBrangefromaminimumof-66(correspondingtoacodeof0forR0)toamaximumof-4(correspondingtoacodeof31forR0).Thestepsizeofthequantizeris2(2dBsteps).R0ischosenas:R0whichminimizesabs(R0-(RdB+66)/2)(29)whereR0cantakeontheintegervaluesfrom0to31correspondingtothe32codesforR0.DecodingoftheR0codeisgivenby:()()()RRR010206610=−max/(30)4.1.6SoftinterpolationofthespectralparametersInterpolationoftheshorttermfilterparametersimprovestheperformanceoftheGSMhalfrateencoder.Thedirectformfiltercoefficients(αi's),whichcorrespondtoquantizedreflectioncoefficients,arethespectralparametersusedforinterpolation.TheGSMhalfratespeechencoderuseseitheraninterpolatedsetofαi'soranuninterpolatedsetofαi's,choosingthesetwhichgivesbetterpredictiongainfortheframe.TwosetsofLPCcoefficientvectorsaregenerated:thefirstcorrespondstotheinterpolatedcoefficients,thesecondtotheuninterpolatedcoefficients.Theframe'sspeechsamplesareinversefilteredusingeachofthetwocoefficientsets,andtheresidualframeenergycorrespondingtoeachsetiscomputed.Thecoefficientsetyieldingthelowerframeresidualenergyisthenselectedtobeused.Iftheresidualenergiesareequal,theuninterpolatedcoefficientsetisused.INT_LPC,asoftinterpolationbit,issetto1wheninterpolationisselectedorto0otherwise.Togeneratetheinterpolatedcoefficientset,thecoderinterpolatestheαi'sforthefirst,second,andthirdsubframesofeachframe.Thefourthsubframeusestheuninterpolatedαi'sforthatframe.Theinterpolationisdoneasfollows.Letαi,Lbethedirect-formLPCcoefficientscorrespondingtothelastframe,αi,Cbethedirect-formLPCcoefficientscorrespondingtothecurrentframe,andDeltobetheinterpolationcurveused.Theinterpolateddirect-formLPCcoefficientvectoratthej-thsubframeofthecurrentframe,αi,j,isgivenby:SIST EN 300 969 V7.0.1:2003

ETSIETSIEN300969V7.0.1(2000-01)18(GSM06.20version7.0.1Release1998)ααααijiLiciLDeljINTSOFT,,,,(,_)()=+−,1≤i≤Np,1≤j≤4(31)ThevaluesoftheinterpolationcurveDelaregivenintable2.Table2:ValuesoftheinterpolationcurveDeljDel(j,0)Del(j,1)10,00,3021,00,6231,00,9241,01,00Fromthispointon,thesubframeindexjisomittedforsimplicitywhenreferringtoαi,jcoefficients,althoughitisimplied.Forinterpolatedsubframes,theαi'sareconvertedtoreflectioncoefficientstocheckforfilterstability.Iftheresultingfilterisunstable,thenuninterpolatedcoefficientsareusedforthatsubframe.Theuninterpolatedcoefficientsusedforsubframe1arethepreviousframe'scoefficients.Theuninterpolatedcoefficientsusedforsubframes2,3,and4arethecurrentframe'scoefficients.4.1.7SpectralnoiseweightingfiltercoefficientsToexploitthenoisemaskingpotentialoftheformants,spectralnoiseweightingisapplied.Thecomputationoftheαicoefficients,usedbyspectralnoiseweightingfiltersW(z)andH(z),isnowdescribed.Defineanimpulsesequenceδ(n)overNssamples:δ(),010=δ(),n=00(32)where1≤n≤Ns-1andh3(n)isthezero-stateresponseofthecascadeofthreefilterstoδ(n).ThethreefiltersareanLPCsynthesisfilter,aninversefilterusingaweightingfactorof0,93andasynthesisfilterwithaweightingfactorof0,7.Inequationform:hnnhniiiNp111()()()=+−=δα0≤n≤Ns-1(33)hnhnhniiiiNp2111093()()(,)()=−−=α0≤n≤Ns-1(34)hnhnhniiiNip321307()()(,)(),=+−=α0≤n≤Ns-1(35)whereαi'sarethedirectformLPcoefficients.Theautocorrelationsequenceofh3(n)iscalculatedusing:RihnhnihniNs3331()()(),=−=−0≤i≤Np(36)FromRih3()thereflectioncoefficientswhichdefinethecombinedspectrallynoiseweightedsynthesisfilterarecomputedusingtheAFLATrecursiononceperframe.SIST EN 300 969 V7.0.1:2003

ETSIETSIEN300969V7.0.1(2000-01)19(GSM06.20version7.0.1Release1998)STEP1DefinetheinitialconditionsfortheAFLATrecursion:PiRih03()(),=0≤i≤Np-1(37)()ViRih031(),=+1-Np≤i≤Np-1(38)STEP2Initializej,theindexofthelatticestage,topointtothefirstlatticestage:j=1STEP3Computerj,thej-threflectioncoefficient,using:rjVPjj=−−−1100()()(39)STEP4Givenrj,updatethevaluesofVjandPjarraysusing:()[]PirPirViVijjjjjj()()()()=+++−−−−12111,0≤i≤Np-j-1(40)()()ViVirVirPijjjjjj()(),=++−−++−−−121111211+j-Np≤i≤Np-j-1(41)STEP5Incrementj:j=j+1STEP6Ifj≤NpgotoSTEP3,otherwiseallNpreflectioncoefficientshavebeenobtained.STEP7Thereflectioncoefficients,rj,arethenconvertedtodirect-formLPCfiltercoefficients,αiforimplementingthecombinedspectrallynoiseweightedsynthesisfilterH(z)andthefilterW(z).ThemethodforthespectralnoiseweightingfiltercoefficientupdatemimickshowthedirectformLPCfiltercoefficientsareupdatedatsubframesofaframe(subclause4.1.6).Nostabilitycheckofinterpolatedspectralnoiseweightingfiltercoefficientsisdoneatsubframes1,2,or3iftheinterpolationflag,INT_LPC="1",butifuninterpolatedcoefficientsareusedatsubframes1,2,and/or3duetoinstabilityoftheunweightedcoefficients(INT_LPC="0"),uninterpolatedweightingfiltercoefficientsarealsousedatthosesubframes.4.1.8LongTermPredictorlagdeterminationFigure3illustratesthatthelongtermlagoptimizationlooksjustlikeacodebooksearchwherethecodebookisdefinedbythelongtermfilterstateandthespecificvectorinthecodebookispointedtobythelongtermpredictorlag,L.Theinputp(n)istheweightedinputspeechforthesubframeminusthezeroinputresponseofjusttheH(z)filter.Σ( )2X-Lββββb(n)Lp'(n)p(n)e(n)totalweightederrorLongtermfilterstateH(z)Figure3:LongtermpredictorlagsearchTheGSMhalfratespeechencoderusesacombinationofopenloopandclosedlooptechniquesinchoosingthelongtermpredictorlag.Firstanopenloopsearchisconductedtodetermine"candidate"lagsateachsubframe.Thenatmost,twobestcandidatelagsateachsubframeareselected,witheachservingasananchorpointforconstructinganopenloopframelagtrajectory,subjecttoamaximumdeltacodingconstraint.TheframelagtrajectorywhichminimizestheSIST EN 300 969 V7.0.1:2003

ETSIETSIEN300969V7.0.1(2000-01)20(GSM06.20version7.0.1Release1998)openloopLTPspectrallyweightederrorenergyfortheframeisthenchosen.TheopenloopLTPpredictiongainscorrespondingtothewinningtrajectoryareusedtoselectthevoicingmode1,2or3.IfMODE≠0,theclosedlooplagevaluationisinitiated.Thewinningtrajectoryhasassociatedwithitalistoflagstobesearchedclosedloopateachsubframe.ItispossibletoallowLtotakeonfractionalvalues,thusincreasingtheresolution,andinturntheperformance,oftheadaptivecodebook.Table3showstheallowablelags.Table3:AllowablelagsRangeResolutionNumberoflagsinrange21to222/31/3623to345/61/67235to492/31/34550to891/21/28090to142153Theresolutionofthelongtermfilterstatemaybeincreasedbyupsamplingandfilteringthestate.Inthisimplementation,anon-causal,zero-phaseFiniteImpulseResponse(FIR)filterisused.Whereneeded,thefuturesamplesforthenon-causalfilteringoperationarereplacedbytheoutputofthepredictor.4.1.8.1OpenlooplongtermsearchinitializationAnopen-looplagsearchisdonetonarrowtherangeoflagsoverwhichaclosed-loopsearchwilleventuallybeperformed.Thefirststepsoftheopen-loopsubframelagsearchareasfollows:STEP1Initializethesubframecounterm=1STEP2Theautocorrelationsequenceofy(n)theinputspeech,s(n),filteredbyW,iscalculatedforallallowableintegerlags,andforafewintegerlagsbelowandabovethelowerandupperlimitsforthecurrentsubframe.()()()()CkmynmNynmNksnNss(,),=+−+−−=−1101LPkLPggminmax−≤≤+−221(42)whereLmin=21andLmax=142.ThevaluePgistheorderofonephaseoftheinterpolatingFIRfilterusedtointerpolatethecorrelations.Theenergyofy(n)forthesubframeiscomputed:()()GkmynmNksnNs(,),=+−−=−2011LPkLPggminmax−≤≤+−221(43)STEP3Thesearrays,C(k,m)andG(k,m),aresearchedfortheintegerlagwhichmaximizesC2(k,m)/G(k,m)whereC(k,m)andG(k,m)needtobegreaterthan0.STEP4Ifavalidmaximumisfoundinstep3,thevaluesforthelag,C,andGcorrespondingtothemaximumareretainedinthearraysasLpeak(0,m),Cpeak(0,m),andGpeak(0,m).Otherwise,Lpeak(0,m)=Lmin(44)Cpeak(0,m)=0(45)Gpeak(0,m)=1(46)STEP5m=m+1SIST EN 300 969 V7.0.1:2003

ETSIETSIEN300969V7.0.1(2000-01)21(GSM06.20version7.0.1Release1998)STEP6Ifm≤4,gotostep2STEP7CalculatetheopenloopframeLTPpredictiongain:PRmRmCmGmvmpeakpeakm=−==1000001014214log(,)(,)(,)(,)(47)where()()RmynmNssnN(,)01201=+−=−,(48)STEP8Determineifthevoicingmodeisunvoiced:IfPv<1,7thenMODE=0,thelongtermpredictorisdisabledandtheopenloopandclosedlooplagsearchesareaborted.Inthiscase,proceedtosubclause4.1.10.4.1.8.2OpenlooplagsearchWhenMODE≠0,thelagsearchprocessingiscontinued.Thenextpartofthesearchfindstheallowablelag(seetable3)whichmaximizesCG2inthevicinityofthebestopen-loopintegerresolutionlag,Lpeak(0,m),forvaluesofC>0.STEP1Initializethesubframecounterm=1STEP2InitializethepeakindexLp,m=0STEP3UsinginterpolatedversionsoftheCandGarrays,allowablelagvaluesk'intherange:Lpeak(0,m)-1
ETSIETSIEN300969V7.0.1(2000-01)22(GSM06.20version7.0.1Release1998)()CkgiCkimIji()(),=−+=305(51)()GkgiGkimIji()(),=−+=305(52)and()jkk=−6(53)Thecoefficientsoftheinterpolatingfilteraregj(i)for0≤i≤5.OnlyCI(k)>0andGI(k)>0valuesareconsidered.Ifnopositivecorrelationisfound,thensetλhnw,m=0,Lpeak(1,m)=Lmin,andgotoStep22.Otherwise,storetheinformationrelatedtothevalidbestallowablelagk.Lp,m=Lp,m+1(54)Lpeak(Lp,m,m)=k(55)Cpeak(Lp,m,m)=CI(k)(56)Gpeak(Lp,m,m)=GI(k)(57)ThenextpartofthesearchevaluatesCG2,forC>0andG>0,atthesubmultiplesofthelagLpeak(Lp,m,m)tofindcandidatepeaks.STEP4InitializethedivisorJ=2STEP5Findnearestintegerlagcorrespondingtosubmultipleofmaximumpeakk1=round[Lpeak(1,m)/J](58)STEP6DetermineifsubmultipleiswithinallowablelagrangeIfk1
ETSIETSIEN300969V7.0.1(2000-01)23(GSM06.20version7.0.1Release1998)STEP8Determineifmaximuminstep7isapeakIf()()()()CkmGkmCkmGkm2211',',',',−−>(60)Gotostep11If()()()()CkmGkmCkmGkm2211',',',',++>(61)Gotostep11STEP9Apeakhasbeenfoundatanintegerlag,k'.UsinginterpolatedversionsoftheCandGarrays,allowablelagvalueswithin+1(exclusive)ofk'aresearched.FindkwhereCkGkII2()()(62)isamaximum,where()CkgiCkimIji()(),=−+=305(63)()GkgiGkimIji()(),=−+=305(64)where()jkk=−6(65)andk'-10andGI(k)>0areconsidered.STEP10Ifthepredictiongainexceedsathreshold,thecorrespondinglag,CI,andGIarestoredintheLpeak(),Cpeak(),andGpeak()arrays;othe
...