ETSI EN 301 704 V7.2.1 (2000-05)

2003-01.Slovenski inštitut za standardizacijo. Razmnoževanje celote ali delov tega standarda ni dovoljeno.Digital cellular telecommunications system (Phase 2+) (GSM); Adaptive Multi-Rate (AMR) speech transcoding (GSM 06.90 version 7.2.1 Release 1998)33.070.50Globalni sistem za mobilno telekomunikacijo (GSM)Global System for Mobile Communication (GSM)ICS:Ta slovenski standard je istoveten z:EN 301 704 Version 7.2.1SIST EN 301 704 V7.2.1:2003en01-december-2003SIST EN 301 704 V7.2.1:2003SLOVENSKI
STANDARD
ETSIEN301704V7.2.1(2000-04)EuropeanStandard(Telecommunicationsseries)Digitalcellulartelecommunicationssystem(Phase2+);AdaptiveMulti-Rate(AMR)speechtranscoding(GSM06.90version7.2.1Release1998)GLOBALSYSTEMFORMOBILECOMMUNICATIONSRSIST EN 301 704 V7.2.1:2003

ETSIETSIEN301704V7.2.1(2000-04)2(GSM06.90version7.2.1Release1998)ReferenceREN/SMG-110690Q7R1KeywordsDigitalcellulartelecommunicationssystem,GlobalSystemforMobilecommunication(GSM),speech,AMRETSI650RoutedesLuciolesF-06921SophiaAntipolisCedex-FRANCETel.:+33492944200Fax:+33493654716SiretN°34862356200017-NAF742CAssociationàbutnonlucratifenregistréeàlaSous-PréfecturedeGrasse(06)N°7803/88ImportantnoticeIndividualcopiesofthepresentdocumentcanbedownloadedfrom:http://www.etsi.orgThepresentdocumentmaybemadeavailableinmorethanoneelectronicversionorinprint.Inanycaseofexistingorperceiveddifferenceincontentsbetweensuchversions,thereferenceversionisthePortableDocumentFormat(PDF).Incaseofdispute,thereferenceshallbetheprintingonETSIprintersofthePDFversionkeptonaspecificnetworkdrivewithinETSISecretariat.Usersofthepresentdocumentshouldbeawarethatthedocumentmaybesubjecttorevisionorchangeofstatus.InformationonthecurrentstatusofthisandotherETSIdocumentsisavailableathttp://www.etsi.org/tb/status/Ifyoufinderrorsinthepresentdocument,sendyourcommentto:editor@etsi.frCopyrightNotificationNopartmaybereproducedexceptasauthorizedbywrittenpermission.Thecopyrightandtheforegoingrestrictionextendtoreproductioninallmedia.©EuropeanTelecommunicationsStandardsInstitute2000.Allrightsreserved.SIST EN 301 704 V7.2.1:2003

ETSIETSIEN301704V7.2.1(2000-04)3(GSM06.90version7.2.1Release1998)ContentsIntellectualPropertyRights.5Foreword.51Scope.62References.63Definitions,symbolsandabbreviations.73.1Definitions.73.2Symbols.93.3Abbreviations.124Outlinedescription.134.1Functionaldescriptionofaudioparts.134.2Preparationofspeechsamples.144.2.1PCMformatconversion.144.3PrinciplesoftheGSMadaptivemulti-ratespeechencoder.144.4PrinciplesoftheGSMadaptivemulti-ratespeechdecoder.164.5Sequenceandsubjectiveimportanceofencodedparameters.175Functionaldescriptionoftheencoder.175.1Pre-processing(allmodes).175.2Linearpredictionanalysisandquantization.175.2.1Windowingandauto-correlationcomputation.185.2.2Levinson-Durbinalgorithm(allmodes).195.2.3LPtoLSPconversion(allmodes).205.2.4LSPtoLPconversion(allmodes).215.2.5QuantizationoftheLSPcoefficients.225.2.6InterpolationoftheLSPs.235.2.7MonitoringresonanceintheLPCspectrum(allmodes).245.3Open-looppitchanalysis.255.4Impulseresponsecomputation(allmodes).285.5Targetsignalcomputation(allmodes).285.6Adaptivecodebook.285.6.1Adaptivecodebooksearch.285.6.2Adaptivecodebookgaincontrol(allmodes).325.7Algebraiccodebook.335.7.1Algebraiccodebookstructure.335.7.2Algebraiccodebooksearch.355.8Quantizationoftheadaptiveandfixedcodebookgains.395.8.1Adaptivecodebookgainlimitationinquantization.395.8.2Quantizationofcodebookgains.395.8.3Updatepastquantizedadaptivecodebookgainbuffer(allmodes).415.9Memoryupdate(allmodes).416Functionaldescriptionofthedecoder.426.1Decodingandspeechsynthesis.426.2Post-processing.456.2.1Adaptivepost-filtering(allmodes).456.2.2High-passfilteringandup-scaling(allmodes).467Detailedbitallocationoftheadaptivemulti-ratecodec.468Homingsequences.518.1Functionaldescription.518.2Definitions.518.3Encoderhoming.528.4Decoderhoming.52SIST EN 301 704 V7.2.1:2003

ETSIETSIEN301704V7.2.1(2000-04)4(GSM06.90version7.2.1Release1998)9Bibliography.56AnnexA(informative):Documentchangehistory.57History.58SIST EN 301 704 V7.2.1:2003

ETSIETSIEN301704V7.2.1(2000-04)5(GSM06.90version7.2.1Release1998)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).Thepresentdocumentdescribesthedetailedmappingfrominputblocksof160speechsamplesin13-bituniformPCMformattoencodedblocksof95,103,118,134,148,159,204,and244bitsandfromencodedblocksof95,103,118,134,148,159,204,and244bitstooutputblocksof160reconstructedspeechsampleswithinthedigitalcellulartelecommunicationssystem.ThecontentsofthepresentdocumentissubjecttocontinuingworkwithinSMGandmaychangefollowingformalSMGapproval.ShouldSMGmodifythecontentsofthepresentdocumentitwillbere-releasedwithanidentifyingchangeofreleasedateandanincreaseinversionnumberasfollows:Version7.x.ywhere:7indicatesRelease1998ofGSMPhase2+.xtheseconddigitisincrementedforallchangesofsubstance,i.e.technicalenhancements,corrections,updates,etc.ythethirddigitisincrementedwheneditorialonlychangeshavebeenincorporatedinthespecification.NationaltranspositiondatesDateofadoptionofthisEN:31March2000DateoflatestannouncementofthisEN(doa):30June2000DateoflatestpublicationofnewNationalStandardorendorsementofthisEN(dop/e):31December2000DateofwithdrawalofanyconflictingNationalStandard(dow):31December2000SIST EN 301 704 V7.2.1:2003

ETSIETSIEN301704V7.2.1(2000-04)6(GSM06.90version7.2.1Release1998)1ScopeThepresentdocumentdescribesthedetailedmappingfrominputblocksof160speechsamplesin13-bituniformPCMformattoencodedblocksof95,103,118,134,148,159,204,and244bitsandfromencodedblocksof95,103,118,134,148,159,204,and244bitstooutputblocksof160reconstructedspeechsamples.Thesamplingrateis8000samples/sleadingtoabitratefortheencodedbitstreamof4.75,5.15,5.90,6.70,7.40,7.95,10.2or12.2kbit/s.Thecodingschemeforthemulti-ratecodingmodesistheso-calledAlgebraicCodeExcitedLinearPredictionCoder,hereafterreferredtoasACELP.Themulti-rateACELPcoderisreferredtoasMR-ACELP.Inthecaseofdiscrepancybetweentherequirementsdescribedinthepresentdocumentandthefixedpointcomputationaldescription(ANSI-Ccode)oftheserequirementscontainedinGSM06.73[6],thedescriptioninGSM06.73[6]willprevail.TheANSI-Ccodeisnotdescribedinthepresentdocument,seeGSM06.73[6]foradescriptionoftheANSI-Ccode.Thetranscodingprocedurespecifiedinthepresentdocumentisapplicablefortheadaptivemulti-ratefullrateandhalfratespeechtrafficchannels(TCH)intheGSMsystem.InGSM06.71[5],areferenceconfigurationforthespeechtransmissionchainoftheGSMadaptivemulti-rate(AMR)systemisshown.Accordingtothisreferenceconfiguration,thespeechencodertakesitsinputasa13-bituniformPCMsignaleitherfromtheaudiopartoftheMobileStationoronthenetworkside,fromthePSTNviaan8-bitA-laworµ-lawto13-bituniformPCMconversion.TheencodedspeechattheoutputofthespeechencoderisdeliveredtoachannelencoderunitwhichisspecifiedinGSM05.03[3].Inthereceivedirection,theinverseoperationstakeplace.2ReferencesThefollowingdocumentscontainprovisionswhich,throughreferenceinthistext,constituteprovisionsofthepresentdocument.• Referencesareeitherspecific(identifiedbydateofpublication,editionnumber,versionnumber,etc.)ornon-specific.• Foraspecificreference,subsequentrevisionsdonotapply.• Foranon-specificreference,thelatestversionapplies.• Anon-specificreferencetoanETSshallalsobetakentorefertolaterversionspublishedasanENwiththesamenumber.• ForthisRelease1998document,referencestoGSMdocumentsareforRelease1998versions(version7.x.y).[1]GSM01.04:"Digitalcellulartelecommunicationssystem(Phase2+);Abbreviationsandacronyms".[2]GSM03.50:"Digitalcellulartelecommunicationssystem(Phase2+);TransmissionplanningaspectsofthespeechserviceintheGSMPublicLandMobileNetwork(PLMN)system".[3]GSM05.03:"Digitalcellulartelecommunicationssystem(Phase2+);Channelcoding".[4]GSM06.94:"Digitalcellulartelecommunicationssystem(Phase2+);VoiceActivityDetection(VAD)forAdaptiveMulti-Ratespeechtrafficchannels".[5]GSM06.71:"Digitalcellulartelecommunicationssystem(Phase2+);AdaptiveMulti-Ratespeechprocessingfunctions;Generaldescription".[6]GSM06.73:"Digitalcellulartelecommunicationssystem(Phase2+);ANSI-CcodefortheAdaptiveMulti-Ratespeechcodec".[7]GSM06.74:"Digitalcellulartelecommunicationssystem(Phase2+);TestsequencesfortheGSMAdaptiveMulti-Ratespeechcodec".SIST EN 301 704 V7.2.1:2003

ETSIETSIEN301704V7.2.1(2000-04)7(GSM06.90version7.2.1Release1998)[8]ITU-TRecommendationG.711(1988):"CodingofanaloguesignalsbypulsecodemodulationPulsecodemodulation(PCM)ofvoicefrequencies".[9]ITU-TRecommendationG.726:"40,32,24,16kbit/sadaptivedifferentialpulsecodemodulation(ADPCM)".3Definitions,symbolsandabbreviations3.1DefinitionsForthepurposesofthepresentdocument,thefollowingtermsanddefinitionsapply.adaptivecodebook:Theadaptivecodebookcontainsexcitationvectorsthatareadaptedforeverysubframe.Theadaptivecodebookisderivedfromthelong-termfilterstate.Thelagvaluecanbeviewedasanindexintotheadaptivecodebook.adaptivepostfilter:Thisfilterisappliedtotheoutputoftheshort-termsynthesisfiltertoenhancetheperceptualqualityofthereconstructedspeech.Intheadaptivemulti-ratecodec,theadaptivepostfilterisacascadeoftwofilters:aformantpostfilterandatiltcompensationfilter.AdaptiveMulti-Rate(AMR)codec:Speechandchannelcodeccapableofoperatingatgrossbit-ratesof11.4kbit/s(“half-rate”)and22.8kbit/s(“full-rate”).Inaddition,thecodecmayoperateatvariouscombinationsofspeechandchannelcoding(codecmode)bit-ratesforeachchannelmode.algebraiccodebook:Afixedcodebookwherealgebraiccodeisusedtopopulatetheexcitationvectors(innovationvectors).Theexcitationcontainsasmallnumberofnonzeropulseswithpredefinedinterlacedsetsofpositions.AMRhandover:HandoverbetweentheFRandHRchannelmodestooptimiseAMRoperation.anti-sparsenessprocessing:Anadaptivepost-processingprocedureappliedtothefixedcodebookvectorinordertoreduceperceptualartifactsfromasparsefixedcodebookvector.channelmode:Half-rateorfull-rateoperation.channelmodeadaptation:Thecontrolandselectionofthe(FRorHR)channelmode.channelrepacking:RepackingofHR(andFR)radiochannelsofagivenradiocelltoachievehighercapacitywithinthecell.closed-looppitchanalysis:Thisistheadaptivecodebooksearch,i.e.,aprocessofestimatingthepitch(lag)valuefromtheweightedinputspeechandthelongtermfilterstate.Intheclosed-loopsearch,thelagissearchedusingerrorminimizationloop(analysis-by-synthesis).Intheadaptivemulti-ratecodec,closed-looppitchsearchisperformedforeverysubframe.codecmode:Foragivenchannelmode,thebitpartitioningbetweenthespeechandchannelcodecs.codecmodeadaptation:Thecontrolandselectionofthecodecmodebit-rates.Normally,impliesnochangetothechannelmode.directformcoefficients:Oneoftheformatsforstoringtheshorttermfilterparameters.Intheadaptivemulti-ratecodec,allfilterswhichareusedtomodifyspeechsamplesusedirectformcoefficients.fixedcodebook:Thefixedcodebookcontainsexcitationvectorsforspeechsynthesisfilters.Thecontentsofthecodebookarenon-adaptive(i.e.,fixed).Intheadaptivemulti-ratecodec,thefixedcodebookisimplementedusinganalgebraiccodebook.fractionallags:Asetoflagvalueshavingsub-sampleresolution.Intheadaptivemulti-ratecodecasub-sampleresolutionof1/6thor1/3rdofasampleisused.full-rate(FR):Full-ratechannelorchannelmode.frame:Atimeintervalequalto20ms(160samplesatan8kHzsamplingrate).SIST EN 301 704 V7.2.1:2003

ETSIETSIEN301704V7.2.1(2000-04)8(GSM06.90version7.2.1Release1998)grossbit-rate:Thebit-rateofthechannelmodeselected(22.8kbsor11.4kbs).half-rate(HR):Half-ratechannelorchannelmode.in-bandsignalling:SignallingforDTX,LinkControl,Channelandcodecmodemodification,etc.carriedwithinthetrafficchannel.integerlags:Asetoflagvalueshavingwholesampleresolution.interpolatingfilter:AnFIRfilterusedtoproduceanestimateofsubsampleresolutionsamples,givenaninputsampledwithintegersampleresolution.inversefilter:Thisfilterremovestheshorttermcorrelationfromthespeechsignal.Thefiltermodelsaninversefrequencyresponseofthevocaltract.lag:Thelongtermfilterdelay.Thisistypicallythetruepitchperiod,oritsmultipleorsub-multiple.LineSpectralFrequencies:(seeLineSpectralPair).LineSpectralPair:TransformationofLPCparameters.LineSpectralPairsareobtainedbydecomposingtheinversefiltertransferfunctionA(z)toasetoftwotransferfunctions,onehavingevensymmetryandtheotherhavingoddsymmetry.TheLineSpectralPairs(alsocalledasLineSpectralFrequencies)aretherootsofthesepolynomialsonthez-unitcircle.LPanalysiswindow:Foreachframe,theshorttermfiltercoefficientsarecomputedusingthehighpassfilteredspeechsampleswithintheanalysiswindow.Intheadaptivemulti-ratecodec,thelengthoftheanalysiswindowisalways240samples.Foreachframe,twoasymmetricwindowsareusedtogeneratetwosetsofLPcoefficientinthe12,2kbit/smode.Fortheothermodes,onlyasingleasymmetricwindowisusedtogenerateasinglesetofLPcoefficients.Inthe12,2kbit/smode,nosamplesofthefutureframesareused(nolookahead).Theothermodesusea5mslookahead.LPcoefficients:LinearPrediction(LP)coefficients(alsoreferredasLinearPredictiveCoding(LPC)coefficients)isagenericdescriptivetermfortheshorttermfiltercoefficients.mode:Whenusedalone,referstothesourcecodecmode,i.e.,tooneofthesourcecodecsemployedintheAMRcodec.(Seealsocodecmodeandchannelmode.)open-looppitchsearch:Aprocessofestimatingthenearoptimallagdirectlyfromtheweightedspeechinput.Thisisdonetosimplifythepitchanalysisandconfinetheclosed-looppitchsearchtoasmallnumberoflagsaroundtheopen-loopestimatedlags.Intheadaptivemulti-ratecodec,anopen-looppitchsearchisperformedineveryothersubframe.out-of-bandsignalling:SignallingontheGSMcontrolchannelstosupportlinkcontrol.residual:Theoutputsignalresultingfromaninversefilteringoperation.shorttermsynthesisfilter:Thisfilterintroduces,intotheexcitationsignal,shorttermcorrelationwhichmodelstheimpulseresponseofthevocaltract.perceptualweightingfilter:Thisfilterisemployedintheanalysis-by-synthesissearchofthecodebooks.Thefilterexploitsthenoisemaskingpropertiesoftheformants(vocaltractresonances)byweightingtheerrorlessinregionsneartheformantfrequenciesandmoreinregionsawayfromthem.subframe:Atimeintervalequalto5ms(40samplesat8kHzsamplingrate).vectorquantization:Amethodofgroupingseveralparametersintoavectorandquantizingthemsimultaneously.zeroinputresponse:Theoutputofafilterduetopastinputs,i.e.duetothepresentstateofthefilter,giventhataninputofzerosisapplied.zerostateresponse:Theoutputofafilterduetothepresentinput,giventhatnopastinputshavebeenapplied,i.e.,giventhatthestateinformationinthefilterisallzeroes.SIST EN 301 704 V7.2.1:2003

ETSIETSIEN301704V7.2.1(2000-04)9(GSM06.90version7.2.1Release1998)3.2SymbolsForthepurposesofthepresentdocument,thefollowingsymbolsapply:()AzTheinversefilterwithunquantizedcoefficients()AzTheinversefilterwithquantizedcoefficients()()HzAz=1ThespeechsynthesisfilterwithquantizedcoefficientsaiTheunquantizedlinearpredictionparameters(directformcoefficients)aiThequantifiedlinearpredictionparametersmTheorderoftheLPmodel1Bz()Thelong-termsynthesisfilter()WzTheperceptualweightingfilter(unquantizedcoefficients)γγ12,TheperceptualweightingfactorsFzE()Adaptivepre-filterTTheintegerpitchlagnearesttotheclosed-loopfractionalpitchlagofthesubframeβTheadaptivepre-filtercoefficient(thequantifiedpitchgain)HzAzAzfnd()(/)(/)=γγTheformantpostfilterγnControlcoefficientfortheamountoftheformantpost-filteringγdControlcoefficientfortheamountoftheformantpost-filtering()HztTiltcompensationfilterγtControlcoefficientfortheamountofthetiltcompensationfilteringµγ=tk1'Atiltfactor,withk1'beingthefirstreflectioncoefficient()hnfThetruncatedimpulseresponseoftheformantpostfilterLhThelengthof()hnfrih()Theauto-correlationsof()hnf()AznγTheinversefilter(numerator)partoftheformantpostfilter()1AzdγThesynthesisfilter(denominator)partoftheformantpostfilter()rnTheresidualsignaloftheinversefilter()Aznγ()hntImpulseresponseofthetiltcompensationfilterβscn()TheAGC-controlledgainscalingfactoroftheadaptivepostfilterαTheAGCfactoroftheadaptivepostfilter()Hzh1Pre-processinghigh-passfilterwnI(),wnII()LPanalysiswindowsLI1()LengthofthefirstpartoftheLPanalysiswindowwnI()LI2()LengthofthesecondpartoftheLPanalysiswindowwnI()LII1()LengthofthefirstpartoftheLPanalysiswindowwnII()LII2()LengthofthesecondpartoftheLPanalysiswindowwnII()rkac()Theauto-correlationsofthewindowedspeechsn'()()wilagLagwindowfortheauto-correlations(60Hzbandwidthexpansion)SIST EN 301 704 V7.2.1:2003

ETSIETSIEN301704V7.2.1(2000-04)10(GSM06.90version7.2.1Release1998)f0ThebandwidthexpansioninHzfsThesamplingfrequencyinHzrkac'()Themodified(bandwidthexpanded)auto-correlations()EiLDThepredictionerrorintheithiterationoftheLevinsonalgorithmkiTheithreflectioncoefficientaji()ThejthdirectformcoefficientintheithiterationoftheLevinsonalgorithm()′Fz1SymmetricLSFpolynomial()′Fz2AntisymmetricLSFpolynomial()Fz1Polynomial()′Fz1withrootz=−1eliminated()Fz2Polynomial()′Fz2withrootz=1eliminatedqiThelinespectralpairs(LSPs)inthecosinedomainqAnLSPvectorinthecosinedomain()qinThequantifiedLSPvectorattheithsubframeoftheframenωiThelinespectralfrequencies(LSFs)Txm()AmthorderChebyshevpolynomialfifi12(),()ThecoefficientsofthepolynomialsFz1()andFz2()fifi12''(),()Thecoefficientsofthepolynomials()′Fz1and()′Fz2fi()Thecoefficientsofeither()Fz1or()Fz2()CxSumpolynomialoftheChebyshevpolynomialsxCosineofangularfrequencyωλkRecursioncoefficientsfortheChebyshevpolynomialevaluationfiThelinespectralfrequencies(LSFs)inHz[]ftfff=1210ThevectorrepresentationoftheLSFsinHz()z()1n,()z()2nThemean-removedLSFvectorsatframen()r()1n,()r()2nTheLSFpredictionresidualvectorsatframenp()nThepredictedLSFvectoratframen()()r21n−ThequantifiedsecondresidualvectoratthepastframefkThequantifiedLSFvectoratquantizationindexkELSPTheLSPquantizationerrorwii,,,,=110LSP-quantizationweightingfactorsdiThedistancebetweenthelinespectralfrequenciesfi+1andfi−1()hnTheimpulseresponseoftheweightedsynthesisfilterOkThecorrelationmaximumofopen-looppitchanalysisatdelaykOiti,,,=13Thecorrelationmaximaatdelaystii,,,=13()Mtiii,,,,=13ThenormalizedcorrelationmaximaMiandthecorrespondingdelaystii,,,=13HzWzAzAzAz()()(/)()(/)=γγ12Theweightedsynthesisfilter()Azγ1Thenumeratoroftheperceptualweightingfilter()12AzγThedenominatoroftheperceptualweightingfilterSIST EN 301 704 V7.2.1:2003

ETSIETSIEN301704V7.2.1(2000-04)11(GSM06.90version7.2.1Release1998)T1Theintegernearesttothefractionalpitchlagoftheprevious(1stor3rd)subframesn'()Thewindowedspeechsignal()snwTheweightedspeechsignal()snReconstructedspeechsignal()′snThegain-scaledpost-filteredsignal()snfPost-filteredspeechsignal(beforescaling)()xnThetargetsignalforadaptivecodebooksearch()xn2,x2tThetargetsignalforalgebraiccodebooksearchresnLP()TheLPresidualsignal()cnThefixedcodebookvector()vnTheadaptivecodebookvectorynvnhn()=()()∗Thefilteredadaptivecodebookvector()ynkThepastfilteredexcitation()unTheexcitationsignal()unTheemphasizedadaptivecodebookvector'()unThegain-scaledemphasizedexcitationsignalTopThebestopen-looplagtminMinimumlagsearchvaluetmaxMaximumlagsearchvalue()RkCorrelationtermtobemaximizedintheadaptivecodebooksearchb24TheFIRfilterforinterpolatingthenormalizedcorrelationterm()Rk()RktTheinterpolatedvalueof()Rkfortheintegerdelaykandfractiontb60TheFIRfilterforinterpolatingthepastexcitationsignal()untoyieldtheadaptivecodebookvector()vnAkCorrelationtermtobemaximizedinthealgebraiccodebooksearchatindexkCkThecorrelationinthenumeratorofAkatindexkEDkTheenergyinthedenominatorofAkatindexkdHx=t2Thecorrelationbetweenthetargetsignal()xn2andtheimpulseresponse()hn,i.e.,backwardfilteredtargetHThelowertriangularToeplizconvolutionmatrixwithdiagonal()h0andlowerdiagonals()()hh139,,Φ=HHtThematrixofcorrelationsof()hndn()Theelementsofthevectordφ(,)ijTheelementsofthesymmetricmatrixΦckTheinnovationvectorCThecorrelationinthenumeratorofAkmiThepositionoftheithpulseϑiTheamplitudeoftheithpulseNpThenumberofpulsesinthefixedcodebookexcitationEDTheenergyinthedenominatorofAkSIST EN 301 704 V7.2.1:2003

ETSIETSIEN301704V7.2.1(2000-04)12(GSM06.90version7.2.1Release1998)()resnLTPThenormalizedlong-termpredictionresidual()bnThesignalusedforpresettingthesignsinalgebraiccodebooksearch()snbThesignsignalforthealgebraiccodebooksearch()′dnSignextendedbackwardfilteredtargetφ'(,)ijThemodifiedelementsofthematrixΦ,includingsigninformationzt,()znThefixedcodebookvectorconvolvedwith()hn()EnThemean-removedinnovationenergy(indB)EThemeanoftheinnovationenergy()~EnThepredictedenergy[]bbbb1234TheMApredictioncoefficients()RkThequantifiedpredictionerroratsubframekEIThemeaninnovationenergyRn()Thepredictionerrorofthefixed-codebookgainquantizationEQThequantizationerrorofthefixed-codebookgainquantizationen()Thestatesofthesynthesisfilter()1Az()enwTheperceptuallyweightederroroftheanalysis-by-synthesissearchηThegainscalingfactorfortheemphasizedexcitationgcThefixed-codebookgain′gcThepredictedfixed-codebookgaingcThequantifiedfixedcodebookgaingpTheadaptivecodebookgaingpThequantifiedadaptivecodebookgainγgcccgg=′Acorrectionfactorbetweenthegaingcandtheestimatedone′gcγgcTheoptimumvalueforγgcγscGainscalingfactor3.3AbbreviationsForthepurposesofthepresentdocument,thefollowingabbreviationsapply.FurtherGSMrelatedabbreviationsmaybefoundinGSM01.04[1].ACELPAlgebraicCodeExcitedLinearPredictionAGCAdaptiveGainControlAMRAdaptiveMulti-RateCELPCodeExcitedLinearPredictionEFREnhancedFullRateFIRFiniteImpulseResponseFRFullRateHRHalfRateISPPInterleavedSingle-PulsePermutationLPLinearPredictionLPCLinearPredictiveCodingLSFLineSpectralFrequencyLSPLineSpectralPairLTPLongTermPredictor(orLongTermPrediction)MAMovingAverageSIST EN 301 704 V7.2.1:2003

ETSIETSIEN301704V7.2.1(2000-04)13(GSM06.90version7.2.1Release1998)4OutlinedescriptionThepresentdocumentisstructuredasfollows:Section4.1containsafunctionaldescriptionoftheaudiopartsincludingtheA/DandD/Afunctions.Section4.2describestheconversionbetween13-bituniformand8-bitA-laworµ-lawsamples.Sections4.3and4.4presentasimplifieddescriptionoftheprinciplesoftheAMRcodecencodinganddecodingprocessrespectively.Insubclause4.5,thesequenceandsubjectiveimportanceofencodedparametersaregiven.Section5presentsthefunctionaldescriptionoftheAMRcodecencoding,whereasclause6describesthedecodingprocedures.Insection7,thedetailedbitallocationoftheAMRcodecistabulated.4.1FunctionaldescriptionofaudiopartsTheanalogue-to-digitalanddigital-to-analogueconversionwillinprinciplecomprisethefollowingelements:1)AnaloguetouniformdigitalPCM−microphone;−inputleveladjustmentdevice;−inputanti-aliasingfilter;−sample-holddevicesamplingat8kHz;−analogue−to−uniformdigitalconversionto13−bitrepresentation.Theuniformformatshallberepresentedintwo'scomplement.2)UniformdigitalPCMtoanalogue−conversionfrom13−bit/8kHzuniformPCMtoanalogue;−aholddevice;−reconstructionfilterincludingx/sin(x)correction;−outputleveladjustmentdevice;−earphoneorloudspeaker.Intheterminalequipment,theA/Dfunctionmaybeachievedeither−bydirectconversionto13-bituniformPCMformat;−orbyconversionto8-bitA-laworµ-lawcompoundedformat,basedonastandardA-laworµ-lawcodec/filteraccordingtoITU-TRecommendationsG.711[8]andG.714,followedbythe8-bitto13-bitconversionasspecifiedinsubclause4.2.1.FortheD/Aoperation,theinverseoperationstakeplace.InthelattercaseitshouldbenotedthatthespecificationsinITU-TG.714(supersededbyG.712)areconcernedwithPCMequipmentlocatedinthecentralpartsofthenetwork.Whenusedintheterminalequipment,thepresentdocumentdoesnotonitsownensuresufficientout-of-bandattenuation.Thespecificationofout-of-bandsignalsisdefinedinGSM03.50[2]inclause2.SIST EN 301 704 V7.2.1:2003

ETSIETSIEN301704V7.2.1(2000-04)14(GSM06.90version7.2.1Release1998)4.2PreparationofspeechsamplesTheencoderisfedwithdatacomprisingofsampleswitharesolutionof13bitsleftjustifiedina16-bitword.Thethreeleastsignificantbitsaresetto'0'.Thedecoderoutputsdatainthesameformat.Outsidethespeechcodecfurtherprocessingmustbeappliedifthetrafficdataoccursinadifferentrepresentation.4.2.1PCMformatconversionTheconversionbetween8-bitA-Laworµ-lawcompresseddataandlineardatawith13-bitresolutionatthespeechencoderinputshallbeasdefinedinITU-TRec.G.711[8].ITU-TRec.G.711[8]specifiestheA-Laworµ-lawtolinearconversionandviceversabyprovidingtableentries.Examplesonhowtoperformtheconversionbyfixed-pointarithmeticcanbefoundinITU-TRec.G.726[9].Section4.2.1ofG.726[9]describesA-Laworµ-lawtolinearexpansionandsubclause4.2.8ofG.726[9]providesasolutionforlineartoA-Laworµ-lawcompression.4.3PrinciplesoftheGSMadaptivemulti-ratespeechencoderTheAMRcodecuseseightsourcecodecswithbit-ratesof12.2,10.2,7.95,7.40,6.70,5.90,5.15and4.75kbit/s.Thecodecisbasedonthecode-excitedlinearpredictive(CELP)codingmodel.A10thorderlinearprediction(LP),orshort-term,synthesisfilterisusedwhichisgivenby:()()HzAzaziiim==+−=1111,(1)where,,,,aimi=1arethe(quantified)linearprediction(LP)parameters,andm=10isthepredictororder.Thelong-term,orpitch,synthesisfilterisgivenby:()111BzgzpT=−−,(2)whereTisthepitchdelayandgpisthepitchgain.Thepitchsynthesisfilterisimplementedusingtheso-calledadaptivecodebookapproach.TheCELPspeechsynthesismodelisshowninfigure2.Inthismodel,theexcitationsignalattheinputoftheshort-termLPsynthesisfilterisconstructedbyaddingtwoexcitationvectorsfromadaptiveandfixed(innovative)codebooks.Thespeechissynthesizedbyfeedingthetwoproperlychosenvectorsfromthesecodebooksthroughtheshort-termsynthesisfilter.Theoptimumexcitationsequenceinacodebookischosenusingananalysis-by-synthesissearchprocedureinwhichtheerrorbetweentheoriginalandsynthesizedspeechisminimizedaccordingtoaperceptuallyweighteddistortionmeasure.Theperceptualweightingfilterusedintheanalysis-by-synthesissearchtechniqueisgivenby:()()()WzAzAz=γγ12,(3)where()AzistheunquantizedLPfilterand0121<<≤γγaretheperceptualweightingfactors.Thevaluesγ109=.(forthe12.2and10.2kbit/smode)or94.01=γ(forallothermodes)andγ206=.areused.TheweightingfilterusestheunquantizedLPparameters.Thecoderoperatesonspeechframesof20mscorrespondingto160samplesatthesamplingfrequencyof8000sample/s.Ateach160speechsamples,thespeechsignalisanalysedtoextracttheparametersoftheCELPmodel(LPfiltercoefficients,adaptiveandfixedcodebooks'indicesandgains).Theseparametersareencodedandtransmitted.AtSIST EN 301 704 V7.2.1:2003

ETSIETSIEN301704V7.2.1(2000-04)15(GSM06.90version7.2.1Release1998)thedecoder,theseparametersaredecodedandspeechissynthesizedbyfilteringthereconstructedexcitationsignalthroughtheLPsynthesisfilter.Thesignalflowattheencoderisshowninfigure3.LPanalysisisperformedtwiceperframeforthe12.2kbit/smodeandoncefortheothermodes.Forthe12.2kbit/smode,thetwosetsofLPparametersareconvertedtolinespectrumpairs(LSP)andjointlyquantizedusingsplitmatrixquantization(SMQ)with38bits.Fortheothermodes,thesinglesetofLPparametersisconvertedtolinespectrumpairs(LSP)andvectorquantizedusingsplitvectorquantization(SVQ).Thespeechframeisdividedinto4subframesof5mseach(40samples).Theadaptiveandfixedcodebookparametersaretransmittedeverysubframe.ThequantizedandunquantizedLPparametersortheirinterpolatedversionsareuseddependingonthesubframe.Anopen-looppitchlagisestimatedineveryothersubframe(exceptforthe5.15and4.75kbit/smodesforwhichitisdoneonceperframe)basedontheperceptuallyweightedspeechsignal.Thenthefollowingoperationsarerepeatedforeachsubframe:Thetargetsignal()xniscomputedbyfilteringtheLPresidualthroughtheweightedsynthesisfilter()()WzHzwiththeinitialstatesofthefiltershavingbeenupdatedbyfilteringtheerrorbetweenLPresidualandexcitation(thisisequivalenttothecommonapproachofsubtractingthezeroinputresponseoftheweightedsynthesisfilterfromtheweightedspeechsignal).Theimpulseresponse,()hnoftheweightedsynthesisfilteriscomputed.Closed-looppitchanalysisisthenperformed(tofindthepitchlagandgain),usingthetarget()xnandimpulseresponse()hn,bysearchingaroundtheopen-looppitchlag.Fractionalpitchwith1/6thor1/3rdofasampleresolution(dependingonthemode)isused.Thetargetsignal()xnisupdatedbyremovingtheadaptivecodebookcontribution(filteredadaptivecodevector),andthisnewtarget,()xn2,isusedinthefixedalgebraiccodebooksearch(tofindtheoptimuminnovation).Thegainsoftheadaptiveandfixedcodebookarescalarquantifiedwith4and5bitsrespectivelyorvectorquantifiedwith6-7bits(withmovingaverage(MA)predictionappliedtothefixedcodebookgain).Finally,thefiltermemoriesareupdated(usingthedeterminedexcitationsignal)forfindingthetargetsignalinthenextsubframe.ThebitallocationoftheAMRcodecmodesisshownintable1.Ineach20msspeechframe,95,103,118,134,148,159,204or244bitsareproduced,correspondingtoabit-rateof4.75,5.15,5.90,6.70,7.40,7.95,10.2or12.2kbit/s.Moredetailedbitallocationamongthecodecparametersisgivenintables9a-9h.Notethatthemostsignificantbits(MSB)arealwayssentfirst.SIST EN 301 704 V7.2.1:2003

ETSIETSIEN301704V7.2.1(2000-04)16(GSM06.90version7.2.1Release1998)Table1:BitallocationoftheAMRcodingalgorithmfor20msframeModeParameter1stsubframe2ndsubframe3rdsubframe4thsubframetotalperframe2LSPsets3812.2kbit/sPitchdelay969630(GSMEFR)Pitchgain444416Algebraiccode35353535140Codebookgain555520Total244LSPset2610.2kbit/sPitchdelay858526Algebraiccode31313131124Gains777728Total204LSPsets277.95kbit/sPitchdelay868628Pitchgain444416Algebraiccode1717171768Codebookgain555520Total159LSPset267.40kbit/sPitchdelay858526(DAMPSEFR)Algebraiccode1717171768Gains777728Total148LSPset266.70kbit/sPitchdelay848424Algebraiccode1414141456Gains777728Total134LSPset265.90kbit/sPitchdelay848424Algebraiccode1111111144Gains666624Total118LSPset235.15kbit/sPitchdelay844420Algebraiccode999936Gains666624Total103LSPset234.75kbit/sPitchdelay844420Algebraiccode999936Gains8816Total954.4PrinciplesoftheGSMadaptivemulti-ratespeechdecoderThesignalflowatthedecoderisshowninfigure4.Atthedecoder,basedonthechosenmode,thetransmittedindicesareextractedfromthereceivedbitstream.Theindicesaredecodedtoobtainthecoderparametersateachtransmissionframe.TheseparametersaretheLSPvectors,thefractionalpitchlags,theinnovativecodevectors,andthepitchandinnovativegains.TheLSPvectorsareconvertedtotheLPfiltercoefficientsandinterpolatedtoobtainLPfiltersateachsubframe.Then,ateach40-samplesubframe:-theexcitationisconstructedbyaddingtheadaptiveandinnovativecodevectorsscaledbytheirrespectivegains;-thespeechisreconstructedbyfilteringtheexcitationthroughtheLPsynthesisfilter.SIST EN 301 704 V7.2.1:2003

ETSIETSIEN301704V7.2.1(2000-04)17(GSM06.90version7.2.1Release1998)Finally,thereconstructedspeechsignalispassedthroughanadaptivepostfilter.4.5SequenceandsubjectiveimportanceofencodedparametersTheencoderwillproducetheoutputinformationinauniquesequenceandformat,andthedecodermustreceivethesameinformationinthesameway.Intable9a-9h,thesequenceofoutputbitsandthebitallocationforeachparameterisshown.Thedifferentparametersoftheencodedspeechandtheirindividualbitshaveunequalimportancewithrespecttosubjectivequality.Beforebeingsubmittedtothechannelencodingfunctionthebitshavetoberearrangedinthesequenceofimportanceasgivenin05.03[3].5FunctionaldescriptionoftheencoderInthisclause,thedifferentfunctionsoftheencoderrepresentedinfigure3aredescribed.5.1Pre-processing(allmodes)Twopre-processingfunctionsareappliedpriortotheencodingprocess:high-passfilteringandsignaldown-scaling.Down-scalingconsistsofdividingtheinputbyafactorof2toreducethepossibilityofoverflowsinthefixed-pointimplementation.Thehigh-passfilterservesasaprecautionagainstundesiredlowfrequencycomponents.Afilterwithacutofffrequencyof80Hzisused,anditisgivenby:21211911376953.0906005859.11927246903.08544941.1927246093.0)(−−−−+−+−=zzzzzHh.(4)Down-scalingandhigh-passfilteringarecombinedbydividingthecoefficientsatthenumeratorof()Hzh1by2.5.2Linearpredictionanalysisandquantization12.2kbit/smodeShort-termprediction,orlinearprediction(LP),analysisisperformedtwiceperspeechframeusingtheauto-correlationapproachwith30msasymmetricwindows.Nolookaheadisusedintheauto-correlationcomputation.Theauto-correlationsofwindowedspeechareconvertedtotheLPcoefficientsusingtheLevinson-Durbinalgorithm.ThentheLPcoefficientsaretransformedtotheLineSpectralPair(LSP)domainforquantizationandinterpolationpurposes.TheinterpolatedquantifiedandunquantizedfiltercoefficientsareconvertedbacktotheLPfiltercoefficients(toconstructthesynthesisandweightingfiltersateachsubframe).10.2,7.95,7.40,6.70,5.90,5.15,4.75kbit/smodesShort-termprediction,orlinearprediction(LP),analysisisperformedonceperspeechframeusingtheauto-correlationapproachwith30msasymmetricwindows.Alookaheadof40samples(5ms)isusedintheauto-correlationcomputation.Theauto-correlationsofwindowedspeechareconvertedtotheLPcoefficientsusingtheLevinson-Durbinalgorithm.ThentheLPcoefficientsaretransformedtotheLineSpectralPair(LSP)domainforquantizationandinterpolationpurposes.TheinterpolatedquantifiedandunquantizedfiltercoefficientsareconvertedbacktotheLPfiltercoefficients(toconstructthesynthesisandweightingfiltersateachsubframe).SIST EN 301 704 V7.2.1:2003

ETSIETSIEN301704V7.2.1(2000-04)18(GSM06.90version7.2.1Release1998)5.2.1Windowingandauto-correlationcomputation12.2kbit/smodeLPanalysisisperformedtwiceperframeusingtwodifferentasymmetricwindows.ThefirstwindowhasitsweightconcentratedatthesecondsubframeanditconsistsoftwohalvesofHammingwindowswithdifferentsizes.Thewindowisgivenby:wnnLnLnLLnLLLIIIIIIII().46,,,,.46(),,,.()()()()()()()=−−=−+−−=+−05401010540111112112coscosππ(5)ThevaluesLI1160()=andLI280()=areused.Thesecondwindowhasitsweightconcentratedatthefourthsubframeanditconsistsoftwoparts:thefirstpartishalfaHammingwindowandthesecondpartisaquarterofacosinefunctioncycle.Thewindowisgivenby:wnnLnLnLLnLLLIIIIIIIIIIIIIIII().46,,,,(),,,()()()()()()()=−−=−−−=+−05402210124111112112coscosππ(6)wherethevaluesLII1232()=andLII28()=areused.NotethatbothLPanalysesareperformedonthesamesetofspeechsamples.Thewindowsareappliedto80samplesfrompastspeechframeinadditiontothe160samplesofthepresentspeechframe.Nosamplesfromfutureframesareused(nolookahead).AdiagramofthetwoLPanalysiswindowsisdepictedbelow.20ms5msframe(160samples)subframe(40samples)framen-1framentIw(n)IIw(n)Figure1:LPanalysiswindowsTheauto-correlationsofthewindowedspeech()′=snn,,0239,arecomputedby:SIST EN 301 704 V7.2.1:2003

ETSIETSIEN301704V7.2.1(2000-04)19(GSM06.90version7.2.1Release1998)rksnsnkkacnk()'()'(),,,,=−==239010(7)anda60Hzbandwidthexpansionisusedbylagwindowingtheauto-correlationsusingthewindow:()wififilags=−=exp,,12211002π,(8)wheref060=Hzisthebandwidthexpansionandfs=8000Hzisthesamplingfrequency.Further,rac()0ismultipliedbythewhitenoisecorrectionfactor1.0001whichisequivalenttoaddinganoisefloorat-40dB.10.2,7.95,7.40,6.70,5.90,5.15,4.75kbit/smodesLPanalysisisperformedonceperframeusinganasymmetricwindow.Thewindowhasitsweightconcentratedatthefourthsubframeanditconsistsoftwoparts:thefirstpartishalfaHammingwindowandthesecondpartisaquarterofacosinefunctioncycle.Thewindowisgivenbyequation(6)wherethevalues2001=Land402=Lareused.Theauto-correlationsofthewindowedspeech()′=snn,,0239,arecomputedbyequation(7)anda60Hzbandwidthexpansionisusedbylagwindowingtheauto-correlationsusingthewindowofequation(8).Further,rac()0ismultipliedbythewhitenoisecorrectionfactor1.0001whichisequivalenttoaddinganoisefloorat-40dB.5.2.2Levinson-Durbinalgorithm(allmodes)Themodifiedauto-correlationsrracac'().()0100010=andrkrkwkkacaclag'()()(),,,==110areusedtoobtainthedirectformLPfiltercoefficientsakk,,,,=110bysolvingthesetofequations.()arikriikackac''(),,,.−=−==110110(9)Thesetofequationsin(9)issolvedusingtheLevinson-Durbinalgorithm.Thisalgorithmusesthefollowingrecursion:[]EriakarijEiakjiaakaEikEiLDaciijiacjiLDiiijijiiijiLDiLD()'()'()/()()()()()()()()()()0011011111101101112====−−−==−=+=−−−−=−−−−fortodofortodoendendThefinalsolutionisgivenasaajjj==(),,,10110.TheLPfiltercoefficientsareconvertedtothelinespectralpair(LSP)representationforquantizationandinterpolationpurposes.TheconversionstotheLSPdomainandbacktotheLPfiltercoefficientdomainaredescribedinthenextclause.SIST EN 301 704 V7.2.1:2003

ETSIETSIEN301704V7.2.1(2000-04)20(GSM06.90version7.2.1Release1998)5.2.3LPtoLSPconversion(allmodes)TheLPfiltercoefficientsakk,,,=110,areconvertedtothelinespectralpair(LSP)representationforquantizationandinterpolationpurposes.Fora10thorderLPfilter,theLSPsaredefinedastherootsofthesumanddifferencepolynomials:()()()′=+−−FzAzzAz1111(10)and()()()′=−−−FzAzzAz2111,(11)respectively.Thepolynomial()′Fz1and()′Fz2aresymmetricandanti-symmetric,respectively.Itcanbeproventhatallrootsofthesepolynomialsareontheunitcircleandtheyalternateeachother.()′Fz1hasarootz=−1(ωπ=)and()′Fz2hasarootz=1(ω=0).Toeliminatethesetworoots,wedefinethenewpolynomials:()()()FzFzz1111=′+−(12)and()()()FzFzz2211=′−−(13)Eachpolynomialhas5conjugaterootsontheunitcircle()eji±ω,therefore,thepolynomialscanbewrittenas()()Fzqzzii11213912=−+−−=∏,,,(14)and()()Fzqzzii212241012=−+−−=∏,,,,(15)where()qii=cosωwithω ibeingthelinespectralfrequencies(LSF)andtheysatisfytheorderingproperty01210<<<<<ωωωπ.WerefertoqiastheLSPsinthecosinedomain.Sincebothpolynomials()Fz1and()Fz2aresymmetriconlythefirst5coefficientsofeachpolynomialneedtobecomputed.Thecoefficientsofthesepolynomialsarefoundbytherecursiverelations(fori=0to4):()()()()fiaafifiaafiimiimi11121211+=+−+=−++−+−(16)wherem=10isthepredictororder.TheLSPsarefoundbyevaluatingthepolynomials()Fz1and()Fz2at60pointsequallyspacedbetween0andπ andcheckingforsignchanges.Asignchangesignifiestheexistenceofarootandthesignchangeintervalisthendivided4timestobettertracktheroot.TheChebyshevpolynomialsareusedtoevaluate()Fz1and()Fz2.Inthismethodtherootsarefounddirectlyinthecosinedomain{}qi.Thepolynomials()Fz1or()Fz2evaluatedatzej=ωcanbewrittenas:()()FeCxjωω=−25,SIST EN 301 704 V7.2.1:2003

ETSIETSIEN301704V7.2.1(2000-04)21(GSM06.90version7.2.1Release1998)with:()()()()()()()()()()()CxTxfTxfTxfTxfTxf=+++++54321123452,(17)where()()Txmm=cosωisthemthorderChebyshevpolynomial,and()fii,,,=15arethecoefficientsofeither()Fz1or()Fz2,computedusingtheequationsin(16).Thepolynomial()Cxisevaluatedatacertainvalueof()x=cosωusingtherecursiverelation:fordowntoendkxfkCxxfkkk=−+−=−+=++4125521212λλλλλ()()()/,withinitialvaluesλ51=andλ60=.ThedetailsoftheChebyshevpolynomialevaluationmethodarefoundinP.KabalandR.P.Ramachandran[6].5.2.4LSPtoLPconversion(allmodes)OncetheLSPsarequantifiedandinterpolated,theyareconvertedbacktotheLPcoefficientdomain{}ak.TheconversiontotheLPdomainisdoneasfollows.Thecoefficientsof()Fz1or()Fz2arefoundbyexpandingequations(14)and(15)knowingthequantifiedandinterpolatedLSPsqii,=,,110.Thefollowingrecursiverelationisusedtocompute()fi1:()()()()()()()fortofordowntoendendifiqfifijifjfjqfjfjii==−−+−=−=−−+−−−1521221121212111112111withinitialvalues()f101=and()f110−=.Thecoefficients()fi2arecomputedsimilarlybyreplacingqi21−byqi2.Oncethecoefficients()fi1and()fi2arefound,()Fz1and()Fz2aremultipliedby11+−zand11−−z,respectively,toobtain()′Fz1and()′Fz2;thatis:()()()()()()′=+−=′=−−=fififiifififii111222115115,,,,,,.(18)FinallytheLPcoefficientsarefoundby:()()()(
...