USER’S MANUAL
Section 6: AUDIO CODING REFERENCE 115
qualityindistinguishablefromtheoriginal,withnoaudiotestitemfallingbelowthe
“perceptible,butnotannoying”thresholdincontrolledlisteningtests.
TheMPEGtestitemsincludethemostdifficultaudioknowntocodecresearchers,sothiswas
dauntingchallenge.Thethinkingwasthatifacodeccouldpassthistest,itwouldsurelybe
transparentfornormalpr
ogrammateriallikevoiceandpopmusic,whicharemucheasierto
encode.AACdesignerschosetouseanewmodularapproachfortheproject,withcomponents
beingplugged‐intoageneralframeworkinordertomatchspecificapplicationrequirements
andthealways‐presentpe
rformance/complexity/delaytradeoffs.
Comparedtothepreviouslayers,AACtakesadvantageofsuchnewtoolsastemporalnoise
shaping,backwardadaptivelinearpredictionandenhancedjointstereocodingtechniques.AAC
supportsawiderangeofsamplingrates(8–96kHz),bitrates(16–576kbps)andfromoneto48
audiochannels.
AACisalotmoresophisticatedtha
nthepreviousMPEGlayers2&3,providingsignificantlymore
codingpower.Becauseitusesamodularapproach,newideascanbedevelopedandpluggedin
tothebasicstructure.Thishadtheadditionaladvantagethatitwaspossibletocombinethe
variouscomponentsfromdifferentdev
elopers,takingthebestpiecesfromeach.AACwasbuilt
onasimilarstructuretoLayer‐3,andthusretainsmostofitsfeatures.Nevertheless,compared
tothepreviousMPEGlayers,AACbenefitsfromsomeimportantnewadditionstothecoding
toolkit:
• Animprovedfilterbankwithafr
equencyresolutionof2048spectral
components,nearlyfourtimesthenumberusedbyLayer‐3.
• TemporalNoiseShaping,anewandpowerfulelementthatminimizesthe
effectoftemporalspread.Thisbenefitsvoicesignals,inparticular.
• APredictionModuleguidesthequantizertoveryeffectivecodingwhen
thereisanoticeablesignalpattern,li
kehightonality.
• PerceptualNoiseShapingallowsafinercontrolofquantizationresolution,
sobitscanbeusedmoreefficiently.
Becauseitusesthismodularapproach,animplementermaypickandchooseamongthe
componenttoolstomakeaproductwithappropriateperformance/complexityratios.Or,ne
w
modulescanbedevelopedlaterand"pluggedin"toitsbasicstructure.Threedefaultprofiles
havebeendefined,usingdifferentcombinationsoftheavailabletools:
• MainProfile.Usesalltoolsexceptthegaincontrolmodule.Providesthe
highestqualityforapplicationswheretheamountofrandomaccessory
memory(RA
M)neededisnotconstrained.
• Low‐complexityProfile. Deletesthepredictiontoolandreducesthe
temporalnoise‐shapingtoolincomplexity.
• Sample‐rateScaleable(SRS)Profile.Addsthegaincontroltooltothelow
complexityprofile.Allowstheleastcomplexdecoder.